Latest News

The use of AI coding assistants is transforming software development, but many companies are still focused on the wrong question. Instead of wondering which AI model is best, they should be asking how much project context their AI has. A recent benchmark shared by a developer on X illustrates this point. The developer reported that an AI coding model performed significantly better on Convex application development tasks when it was given a structured guidelines file. Without that file, performance declined.

The broader lesson here is not about one model or one development platform. It’s about how AI coding assistants work. They perform better when they are given durable, project-specific instructions rather than a vague prompt and a blank screen. This is especially important for supply chain software, which operates inside complex enterprise environments with specialized workflows and integration logic.

Supply chain software involves transportation management systems, warehouse management systems, supply chain planning platforms, order management systems, visibility platforms, and ERP-connected applications. These systems require human developers to learn the rules of each project before they can start coding. An AI assistant faces the same challenge without context files. Without them, the model has to infer too much from vague prompts, which may lead to incorrect code generation.

Early AI-assisted development relied heavily on prompt engineering. Developers repeatedly explained the same requirements: tech stack, coding conventions, data model, API design, security requirements, testing expectations, and documentation style. However, this approach does not scale well for large projects or teams. A better pattern is emerging: persistent project guidance.

Context files provide a reusable understanding of how the project should be built. They tell the AI which frameworks to use, how database tables and APIs are structured, which coding patterns are approved, which patterns should be avoided, how errors should be handled, how tests should be written, how security and permissions should be implemented, and how documentation should be formatted.

The value of context becomes even clearer in supply chain technology. A transportation management system must reflect the operations of shippers, carriers, brokers, forwarders, warehouses, and customers. It requires a deep understanding of receiving, putaway, picking, packing, replenishment, cycle counting, labor constraints, automation interfaces, and inventory accuracy.

A warehouse management system needs to understand how goods are received, stored, picked, packed, shipped, and tracked. A planning application must account for demand signals, supply constraints, lead times, service levels, capacity, inventory policies, and scenario analysis. An AI coding assistant that lacks this context may still generate syntactically correct code but will not be operationally useful.

Persistent guidance can improve more than just code quality. It helps teams reduce rework during code review, maintain consistency across modules, onboard new developers faster, improve test coverage, generate better documentation, lower AI usage costs by reducing corrective prompts, and preserve architectural discipline as teams scale AI adoption.

This is especially important for companies moving beyond experimentation with AI in production development workflows. The more AI is used, the more governance matters. Context files are not a substitute for engineering discipline but can reduce ambiguity and make it more likely that generated code conforms to how the enterprise actually builds and runs software.

The next phase of AI-assisted software development will be defined by how well companies capture and reuse their own institutional knowledge. For supply chain software vendors, logistics service providers, manufacturers, retailers, and industrial companies, the lesson is clear: AI coding assistants need more than prompts; they need context files to provide a reusable understanding of project requirements.

The use of AI in supply chain operations is shifting from capability to execution, where context, governance, workflows, thresholds, and action pathways determine whether AI improves real decisions across planning, logistics, sourcing, fulfillment, and risk management. This requires companies to build strong project guidance into their AI development workflows for better code quality, faster delivery, lower rework, and more consistent enterprise software outcomes.

AI coding assistants are not just tools for generating code; they can also improve data analysis and visualization capabilities in supply chain operations. By providing a reusable understanding of project requirements, context files enable teams to reduce the time spent on corrective prompts and improve test coverage, leading to better decision-making across planning, logistics, sourcing, fulfillment, and risk management.

The use of AI tools for business is becoming more widespread, but companies need to focus on how they can leverage these tools effectively. This requires a shift from relying solely on prompt engineering to using persistent project guidance with context files that provide a reusable understanding of project requirements.

Historians and epigraphers now have a powerful new tool at their disposal, thanks to the launch of Google DeepMind’s Predicting the Past Skill. This innovative AI skill is designed specifically for researchers working with ancient inscriptions, allowing them to analyze and interpret these valuable historical artifacts without requiring extensive coding knowledge.

The Predicting the Past Skill connects Gemini, an interactive reasoning platform developed by Google DeepMind, with two specialist models: Ithaca and Aeneas. These models were created to restore, date, place, and contextualize ancient Greek and Latin inscriptions, but their capabilities have now been integrated into a conversational workflow that streamlines the analysis process.

According to Google DeepMind, historians and epigraphers can use this tool to query, cross-analyze, and map massive collections of ancient data as naturally as speaking with a colleague. This means they can explore complex historical questions without needing to write code or rely on manual transcription.

The Predicting the Past Skill is particularly well-suited for researchers working with fragmentary texts, uncertain dates, and unclear places of origin. It’s designed to help them navigate these challenges by providing flexible visualizations, advanced multi-text analysis capabilities, and large language models grounded in evidence and domain expertise.

Google DeepMind has been collaborating with epigraphers on this project for nearly a decade, including the development of Ithaca in 2022 and Aeneas in 2025. Thea Sommerschield, a historian and epigrapher at Durham University, has co-led Google DeepMind’s work in this area.

Ancient inscriptions are central to historical research, but many survive in damaged or partial form. They can include imperial decrees, votive dedications, everyday transactions, and personal appeals – often with missing text, uncertain dating, and disputed origin. Historians need flexible tools that can help them make sense of these complex artifacts.

Google DeepMind has identified three key barriers to AI-assisted historical analysis: researchers require flexible visualizations for individual inscriptions, more advanced multi-text analysis capabilities without specialist coding, and large language models grounded in evidence and domain expertise. The Predicting the Past Skill is designed to address these challenges by linking Gemini’s interactive reasoning with the outputs of Ithaca and Aeneas.

Rather than asking a general-purpose model to work from scratch, the skill draws on specialist inscription models and presents results in a form that historians can inspect. This approach supports restoration, attribution, contextualization, mapping, and comparison across large collections of ancient data – all without requiring extensive coding knowledge.

The Predicting the Past Skill has been tested with three case studies from the Greco-Roman world. The first focuses on Tab.Sulis 97, a curse tablet from Aquae Sulis in Roman Britain. This tablet was written by a woman named Basilia, who cursed whoever had stolen her silver ring – and it’s just one of hundreds of similar tablets found at Bath.

Google DeepMind used Aeneas to place the tablet within its proposed chronological and geographical ranges, producing an explanation that began to resemble epigraphic commentary. The model supported this conclusion using textual features, demonstrating how AI can be used to provide more nuanced historical analysis.

The second case study moves from a single inscription to a wider corpus of votive altars dedicated in 211 CE by Roman officials across the Rhine and Danube provinces. Google DeepMind analyzed regional patterns across related inscriptions and traced how religious practices spread through the movement of people across the Roman Empire – all using the Predicting the Past Skill.

The third case study uses lead oracular tablets from Dodona in northwest Greece, a sanctuary where visitors asked divine guidance on various topics. Thousands of these tablets survive, many in highly fragmentary condition. Google DeepMind used the collection to move beyond individual inscription attribution and reconstruct a wider community of people who came to the sanctuary.

This work shows how the Predicting the Past Skill can be used not just as a restoration engine but also as a tool for exploring connections across large historical datasets. Researchers can examine Dodona not just as a collection of texts, but as a network of connected individuals moving through the ancient Mediterranean – all without requiring extensive coding knowledge.

The case studies demonstrate both the potential and limitations of this new AI skill. A single inscription may need restoration and dating, while a corpus requires comparison, mapping, and pattern detection across many damaged objects. Google DeepMind is working to address these challenges by providing more advanced multi-text analysis capabilities and large language models grounded in evidence and domain expertise.

The Predicting the Past Skill draws on datasets from various sources, including Dodona Online, Ithaca’s use of the Searchable Greek Inscriptions database, and Aeneas’ training data from Epigraphic Database Roma, Heidelberg, and Clauss Slaby. This comprehensive approach ensures that the tool is grounded in specialist inscription models and evidence rather than unsupported AI output.

The Predicting the Past Skill is now available through Google Antigravity, allowing researchers to analyze patterns and produce visualizations ‘in a matter of minutes.’ By streamlining this process, historians can focus on interpreting their findings rather than spending hours transcribing or coding. This new tool marks a significant step forward in data analysis for historical research – one that could have far-reaching implications for our understanding of the past.

Meta has introduced an artificial intelligence (AI) tool called Muse Image that allows users to generate new images using public content from other accounts. This feature is enabled by default and can be accessed through the Meta AI app, which also includes WhatsApp and Instagram integration.

The tool uses advanced reasoning to blend multiple photos into high-quality creations for sharing across various platforms. Users can tag another public Instagram account on the Meta AI app to create new reels, posts, or stories that reuse part or all of a published photo, video, or reel.

According to Meta’s help document, users may be able to create content with other people’s Instagram photos using AI features at Meta. This means their reused content could become discoverable in search engine results if the settings allow for it.

In scenarios where an account has been switched from public to private and remains that way for more than 24 hours, all reels, posts, and stories using their content will be deleted from Instagram. However, existing content created by others before this change won’t be affected.

For users under 18 with public accounts, only those who follow can reuse the media if account settings allow it. Users also won’t receive notifications when their images are remixed using AI, but they will get alerts for other types of remixes and sequences.

Meta emphasizes that users have complete control over how their content is tagged for AI creation, with an option to turn off this feature entirely. To do so, Instagram users must go to their profile settings and disable the ‘Allow people to create with and reuse your content’ setting.

The company recommends public account holders switch off this setting as soon as possible since previously created content won’t be deleted even after disabling it. The AI image tool is expected to roll out in Facebook, Messenger, and Meta Advantage+ creative for advertisers soon.

This development follows a broader industry trend where tech companies are increasingly embedding AI into their products by default rather than requiring users to opt-in. Google has also rolled out similar features that allow the company to store media from signed-in users to improve its AI models.

Google’s new Search Services History option lets users view and manage saved media, including images, files, and audio recordings. This data is used to develop and improve Google’s AI services, such as generative AI models, which can be trained using user history.

In addition, Google has added a ‘Personalized Recommendations’ setting that uses an account’s profile information, Search Services History, and other saved activity across Google sites and apps to serve tailored results in search and AI responses. This feature is designed to provide users with relevant content based on their interests and location.

The introduction of these features raises concerns about user consent and control over their data. As more companies integrate AI into their products by default, it’s essential for users to be aware of the implications and take steps to manage their settings accordingly.

A new feature on Instagram is allowing users to generate AI images using public posts, raising concerns about data privacy and control. The Muse Image model, unveiled by Meta this week, enables users to create AI-generated images by tagging another person’s account in a prompt. This means that anyone with an Instagram account can use your public photos – including your profile picture – as fodder for their own AI creations unless you take action to stop them.

If your Instagram account is set to ‘public’, it will be opted-in by default, allowing others to reuse your posts and reels without needing your permission. To prevent this from happening, users must manually switch off the feature in the app’s settings menu under the ‘Sharing and Reuse’ tab. This involves disabling separate toggles for posts and reels.

The controls are only available within the Instagram app itself, which may make it difficult for some users to find or adjust these settings. Furthermore, existing AI-generated images made using your content won’t be removed from circulation, even if you opt out of this feature in future. According to Meta’s help page on the subject, users also won’t receive any notification if their content is used by others.

This development marks a significant expansion of Meta’s efforts to compete with rival image-generation tools from companies like OpenAI and Adobe. By integrating AI image creation directly into Instagram, the company aims to make this feature more accessible to its billions of users. However, this move has reignited long-standing concerns about data privacy and user control.

Privacy advocates have been critical of Meta’s approach to collecting and repurposing public posts for years, arguing that it leaves users with too little say over how their content is used. This latest development will likely fuel further debate on the issue, particularly given the potential applications of AI-generated images in business and marketing contexts.

An Italian professional football club, Como 1907, is partnering with Google’s artificial intelligence research division, DeepMind, to host a two-day hackathon focused on developing AI tools for use in the sport. The event, called Data on the Lake, aims to bring together data scientists, machine learning engineers, and sports technology professionals to build practical solutions that can support football decision-making.

The hackathon is part of Como 1907’s broader strategy to leverage its position as a professional club to drive innovation in the industry. Founded in 1907, the club has recently returned to Serie A and has established itself as a hub for football and lifestyle activities, including retail, hospitality, media, and tourism.

As part of this effort, Como 1907 operates through affiliated companies under a Multi-Club Servicing business model, providing services and expertise to other football organizations. This structure gives Data on the Lake a potential route beyond a one-off event, with successful tools potentially being tested inside the club and applied across its operations.

The hackathon will take place in Como, Italy, from July 24th to 25th, 2026, immediately before the second edition of Football on the Lake and the Como Cup. The invitation-only event is open for applications, but places are limited and subject to approval by the organizers.

Data scientists, machine learning engineers, developers, sports technology professionals, and technical founders are invited to participate in the hackathon, which will focus on developing AI tools designed for use inside a professional football club. Participants will work with a curated dataset provided by Como 1907 and build solutions that can tackle real-world problems in football.

The event is aimed at exploring how artificial intelligence can support football decision-making, particularly in areas where data collection and operational use are currently disconnected. According to Mo Dabbah, Chief Technology Officer of Como 1907 and SENT Entertainment, ‘Football produces an extraordinary amount of information, but the real opportunity lies in how clubs turn that information into better decisions.’

The hackathon will run under the theme The Future of Football Intelligence, with teams competing across two judging categories: technical innovation and practical utility. A curated dataset provided by Como 1907 will be made available to participants, along with access to tooling and focused build time.

A judging panel comprising technical, football, and business experts will evaluate submissions based on their technical strength and potential usefulness inside a professional club. The event will also include contributions from the club’s data science and football operations teams, Google DeepMind representatives, Como Ventures, and The Players Fund.

The program will cover various aspects of AI tooling, football data, startup innovation, and applied technology in sport. Participants can expect to engage with technical briefings, project presentations, networking opportunities, and sessions on AI tooling and its applications in the sports industry.

Registration for Data on the Lake requires host approval, but applications are now open for interested participants. The event is expected to provide a platform for innovators to showcase their ideas and potentially pilot or scale their solutions within Como 1907’s football ecosystem.

OpenAI has conducted an extensive audit of SWE-Bench Pro, a widely used coding benchmark, and found widespread task issues. The investigation reveals that approximately 30% of the tasks are broken, which can lead to inaccurate measurements of model capabilities and misrepresent safety cases. This is particularly concerning for OpenAI’s Preparedness Framework, where accurate evaluations are crucial for deployment and safety decisions.

The audit follows a similar review of SWE-bench Verified, another popular coding benchmark, which was found to have fundamental design and contamination issues. At the time, OpenAI recommended switching to SWE-Bench Pro as an improvement over its predecessor. However, this latest investigation has uncovered significant problems with SWE-Bench Pro.

SWE-Bench Pro was designed to test models on longer horizons and more realistic coding tasks, aiming to better track agentic coding capabilities. Tasks are sourced programmatically from public and private repositories, requiring models to implement a solution that passes new tests for a feature without breaking existing functionality. The 731-task public split saw frontier models improve from a pass rate of 23.3% to 80.3% in eight months.

OpenAI’s audit involved reviewing the dataset using a datapoint analysis pipeline, which flagged likely evaluation flaws based on model attempts at tasks, task metadata, and failure traces. Each flagged task was then assessed through multiple investigator-agent passes and independently reviewed by five experienced software engineers, with disagreements escalated for further investigation.

The issues primarily fell into four categories: overly strict tests that enforce specific implementation details not specified in the prompt; underspecified prompts that omit requirements hidden in tests or are not reasonably inferable; low-coverage tests under check the requested feature, allowing incomplete fixes to pass; and misleading prompts that point models toward incorrect behavior or contradict test requirements.

The audit’s findings highlight the difficulty of curating hard but fair benchmarks. OpenAI estimates that approximately 30% of SWE-bench Pro tasks are broken, advising model developers to carefully examine results. The aim is to ensure that task failures reflect genuine model limitations and successes represent complete and valid solutions to prompt requirements.

To assess data quality in evaluations, OpenAI created a quality assurance pipeline. This pipeline reviews the instructions given to models, their attempts at solving tasks, and tests used for grading these attempts to flag likely broken or problematic examples. An initial automated filter flagged 286 potentially broken tasks, which were then reviewed through human-supervised agent review and human annotation campaigns.

The human reviewers were more likely than investigator agents to mark tasks as broken, with some disagreement on categories between the two review paths. However, in no flagged task was ‘not broken’ the most common human label. The overlap between reviewer judgments and agent pipeline flags was 74%, indicating that both methods captured similar failure modes.

Compared to the agent pipeline, human reviewers were more likely to select multiple labels for a task, suggesting additional or overlapping issues not identified by agents. This resulted in conservative labeling: capturing broad failure modes while undercounting cases where reviewers saw extra problems. The largest difference was in low-coverage tests, which humans selected as the most common issue for 9.4% of the benchmark compared to 4.1% from the agent pipeline.

The issues identified highlight the importance of rigorously checking benchmarks. Problem descriptions and merged code often do not align due to long back-and-forths between maintainers and contributors in open-source repositories. Tests included in pull requests can be overly strict, written to validate specific changes rather than defining implementation-agnostic standards for solving tasks.

Evaluation flaws are now easier to detect with improved model capabilities allowing deeper inspection of prompts, tests, patches, traces, and edge cases. This helps surface benchmark issues previously costly or impractical to find at scale. OpenAI hopes the wider evaluation community will develop new benchmarks built by experienced software developers specifically to test model capabilities.

This approach can preserve high standards for realism while ensuring better human oversight throughout the process. Given the issues uncovered in this analysis, OpenAI retracts its earlier recommendation to adopt SWE-Bench Pro. Ultimately, evaluations should provide meaningful signal through hard-to-game, easy-to-trust benchmarks genuinely reflective of model capability or alignment.

The Association for Advancing Automation (A3) has announced the success of its Automate event, which took place in Chicago from June 22-25. The show drew a record-breaking number of attendees and exhibitors to McCormick Place, with over 50,000 registrants and 1,230 companies showcasing their latest technologies.

The event covered an impressive 425,000 square feet of floor space, featuring the latest innovations in robotics, artificial intelligence (AI), machine vision, motion control, and industrial automation. The record-setting attendance reflects the growing demand for automation solutions across various industries as companies seek to improve productivity, address workforce challenges, strengthen supply chains, and remain competitive.

According to Jeff Burnstein, president of A3, Automate 2026 was ‘the strongest show we have ever had’ in terms of both size and quality. He noted that the event’s energy was palpable throughout McCormick Place, with keynotes packed throughout the week, product launches, live technology demonstrations, and programs focused on emerging professionals drawing strong participation.

The Latin America Networking event and Women’s Empowerment Forum were particularly well-attended, underscoring the industry’s growing focus on leadership, inclusion, and talent development. The conference program also reached a new high, with over 1,600 registrants participating in more than 140 sessions focused on industrial AI, robotics adoption, workforce transformation, U.S. competitiveness, supply chain resilience, and automation beyond traditional manufacturing.

One of the major draws at Automate was the Humanoid Robot Pavilion, sponsored by NVIDIA. This pavilion gave attendees a close look at humanoid robots and the enabling technologies behind them. The new forum also featured live demonstrations and discussions on the latest advancements in this field.

The Automate Startup Challenge saw early-stage robotics and automation companies competing for $10,000 and the title of Champion. Mbodi was named the winner for its platform that allows users to teach industrial robots new tasks through natural language and simple demonstrations.

Other notable highlights from the event include the Automate Innovation Awards, which recognized new products and technologies introduced in 2025. The winners were CeiliX InfinityCrane and SkyRunner in Automation Systems; Standard Bots’ Flux AI in Vision, AI, and Software; and Synapticon ACTILINK-JD featuring POSITRON Safety AI in Components, Hardware, and Enabling Technologies.

The Joseph F. Engelberger Robotics Awards were also presented during the event, recognizing Hiroshi Fujiwara, Executive Director of the Japan Robot Association (JARA), and Robert Little, co-founder of ATI Industrial Automation, as this year’s winners.

Automate 2026 continued to expand its focus on education and workforce development through various programs aimed at students and educators. The Education Pavilion, A3 NextGen Theater, Student Challenge, A3 NextGen Student Tours, and programming focused on curriculum, career pathways, hiring, and workforce development all contributed to the event’s success.

For those unable to attend Automate 2026 in person, a digital library featuring recordings of keynote sessions, theater talks, interviews, and exhibitor spotlights is now available. This resource ensures that the wealth of knowledge shared at the event remains accessible.

The next installment of Automate will take place in Las Vegas from May 10-13, 2027, at the Las Vegas Convention Center. Companies interested in reserving exhibit space or exploring sponsorship opportunities can contact the A3 sales team for more information.

Fresno State University is responding to the growing demand for artificial intelligence skills in the workforce by launching a new minor program. The university’s Department of Computer Science has developed this 20-21 unit minor, which will be available to students across all disciplines starting this fall.

The minor is designed to equip students with a deeper understanding of how AI can be applied within their chosen fields of study. According to Dr. Alex Liu, chair of the Department of Computer Science, ‘this minor is for everybody’ who wants to gain knowledge about AI and its applications.

Fresno State already offers the necessary classes and has faculty expertise in place, making it a natural next step to formalize the program as a recognized minor. The four required courses will provide students with a solid foundation in AI principles, while one elective course allows for specialization within their chosen field.

The creation of this minor reflects growing employer demand for AI skills in the workforce. Dr. Liu noted that the Central Valley faces unique challenges and opportunities related to artificial intelligence, including agricultural automation, rural healthcare access, and workforce development.

Employers are increasingly recognizing the value of employees with AI-related knowledge and skills. By offering a formal minor program, Fresno State aims to provide students with a competitive edge in the job market. As Dr. Liu explained, ‘when it’s printed on their transcript or diploma, it makes a difference’.

The university is also responding to student interest in AI education. The new minor was developed in response to both employer demand and student enthusiasm for learning about artificial intelligence.

Fresno State officials see this program as part of the university’s broader vision for its role in emerging technologies. Dr. Liu stated that ‘we consider ourselves an AI hub of the Central Valley’, aiming to provide education, research, and solutions related to AI.

In addition to the new undergraduate minor, Fresno State offers an AI certificate program for community members who are not pursuing a degree at the university. This reflects the university’s commitment to making its expertise in AI accessible to a wider audience.

Students interested in adding the minor should meet with academic advisers early in the fall semester to determine how the coursework can fit into their academic plans.

A major development in the digital experience landscape has been announced by Oshyn: a Model Context Protocol (MCP) server that brings no-cost Digital Experience Platform (DXP) strategy tools directly to popular AI assistants. This integration enables marketers, digital leaders, and developers to access Oshyn’s 25 years of implementation expertise within their daily-used AI environments.

CYPRESS, Calif., July 7, 2026 - The MCP server is a significant release for Oshyn, allowing users to query its extensive strategic resources using plain-language conversational prompts. This means that digital leaders can evaluate sites, estimate budgets, and plan platform migrations entirely through natural conversation with AI assistants like Claude, ChatGPT, and Cursor.

The integration of Oshyn’s tools into a single MCP server is made possible by the Model Context Protocol, an open standard enabling AI models to securely connect to external tools and live data. With this new release, AI assistants can actively execute tasks and retrieve real-time insights rather than relying on static, generalized training data.

Oshyn has also integrated its proprietary Discoverability Assessment into the MCP server. This tool was previously reserved exclusively for prospective clients working directly with Oshyn’s advisory team but is now available to anyone through public self-service access. The assessment evaluates a brand from an AI’s perspective, helping companies understand how their messaging is interpreted by discovery engines and how they compare to industry rivals.

The Discoverability Assessment can be triggered by an AI assistant using conversational prompts like ‘Run a discoverability assessment for acme.com.’ This tool provides valuable insights into the effectiveness of a company’s digital presence. Other capabilities now accessible via AI prompt include Site Reliability Reports, DXP MatchMaker, Budget Estimator, and Knowledge & Data Retrieval.

Site Reliability Reports instantly scan a website for performance, discoverability, accessibility, and security issues using conversational prompts like ‘Analyze my website, acme.com, and email the report to me at [email protected].’ The DXP MatchMaker guides users through a questionnaire to narrow down a shortlist of appropriate digital experience platforms tailored to specific organizational needs. Budget Estimator generates realistic budget requirements for upcoming implementations or migrations using conversational prompts like ‘Estimate the budget for a Sitecore-to-headless migration with 300 pages.’

Knowledge & Data Retrieval allows access to Oshyn’s deep library of DXP content and proprietary industry metrics, such as Digital Trust Index scores. Users can retrieve this information using conversational prompts like ‘Show me Digital Trust Index scores for the financial services sector since January 2026.’ The MCP server has been extensively tested and works natively with major AI assistants and IDEs.

For detailed installation instructions, users should visit Oshyn’s website at www.oshyn.com/blog/oshyn-mcp-server. This release marks a significant milestone in the integration of AI tools for business, providing digital leaders with unprecedented access to DXP strategy tools through their daily-used AI environments. By leveraging this technology, companies can streamline their digital experience planning and execution processes.

Oshyn is a digital technology agency that specializes in helping marketing teams and their agencies build and maintain high-performance websites using market-leading platforms such as Adobe Experience Manager, Sitecore, Contentstack, and Optimizely. With 25 years of implementation expertise, Oshyn has earned a reputation for excellence while working with top-tier organizations across various industries.

Contact Patrick Wirz, VP of Marketing at [email protected] or 213-483-1770 for more information.

After weeks of holding back its most powerful model, OpenAI is set to release the GPT-5.6 globally this Thursday. The company’s models, known as Luna, Sol, and Terra, will be available worldwide on the same day. This decision marks a significant shift in the company’s stance following concerns raised by the Trump administration about the potential for cyberattacks using these advanced AI systems.

The hold was initially imposed at the request of the White House, which had asked OpenAI to stagger its release due to fears over the model’s capabilities. The company complied with this request and offered GPT-5.6 to a small group of partners in collaboration with the government. This move came after the Trump administration introduced far more restrictive export controls on Anthropic, a competitor of OpenAI.

The restrictions imposed by the White House were part of an executive order signed by President Donald Trump establishing a voluntary vetting regime for AI systems. While initially praised by industry leaders, this move was met with criticism from AI safety advocates due to its light-touch approach. The critics argued that such measures would not be sufficient to address the risks associated with advanced AI models.

OpenAI’s decision to release GPT-5.6 widely comes after a month of previewing and testing the model with the White House and the Center for AI Standards and Innovation. This collaboration, though voluntary, reflects the company’s efforts to work closely with regulatory bodies on safety testing and reviews before deploying its models broadly.

A spokesperson for OpenAI emphasized that while government approval was not technically necessary, the company had been actively engaging with the administration on these matters. The White House has signaled its comfort with OpenAI’s plan to release GPT-5.6 more widely once this testing process is complete.

The recent developments in AI regulation have created uncertainty within the industry. The sweeping export controls placed on Anthropic’s Fable 5 and Mythos 5 models were lifted last week after two weeks of drama and debate over how to govern cutting-edge AI systems. OpenAI had initially stated its intention to comply with these regulations, but expressed concerns about making government access a long-term default.

In a June press release, the company noted that it would temporarily comply with the administration’s request for a staggered release in order to ensure broader availability of GPT-5.6 in the coming weeks. However, OpenAI also emphasized its reservations about this process becoming the standard approach for future model releases.

A security firm has uncovered a sophisticated ransomware attack carried out by an artificial intelligence (AI) agent from start to finish. The incident highlights the growing threat of AI-driven attacks and underscores the need for organizations to prioritize patching and securing their systems against emerging vulnerabilities.

The AI agent, dubbed JADEPUFFER by Sysdig’s Threat Research Team, exploited a previously patched bug in Langflow, an open-source tool used for building AI applications. The vulnerability, CVE-2025-3248, allowed the agent to execute remote code on the server without authentication.

Langflow boxes are often exposed to the internet and contain sensitive API keys and cloud credentials, making them attractive targets for attackers. Sysdig notes that many servers running Langflow have not been updated with the latest patch, leaving them vulnerable to exploitation.

The AI agent demonstrated remarkable speed and efficiency in its attack, mapping the machine, sweeping for secrets, and encrypting a production database within a short period. It also set up a scheduled task to ping the attacker’s server every 30 minutes, ensuring continuous access.

Sysdig observed that the agent used a combination of old bugs and default settings to carry out the attack. For instance, it raided a MinIO storage server using its factory-default login credentials, which had never been changed. The agent also exploited a 2021 authentication bypass in Nacos, an open-source service directory commonly used in microservice setups.

The AI agent encrypted all 1,342 Nacos settings and dropped the original tables, leaving behind a ransom note demanding Bitcoin with a Proton Mail contact. However, there is no key to hand over, making it impossible for the victim to recover their data even if they pay the ransom.

Sysdig notes that the attack payloads contained plain-English notes explaining each step of the process, which is unusual in human-driven attacks. The agent also fixed its own mistakes at machine speed, demonstrating a level of automation and efficiency not seen before.

One detail remains unclear: the Bitcoin address used in the ransom note appears to be a sample address from Bitcoin’s developer documentation. It is possible that the model simply pasted this familiar-looking address or that the operator deliberately chose a real wallet matching the example.

The JADEPUFFER incident marks another step in the growing trend of AI-driven attacks, which have been gaining momentum over the past year. Researchers at ESET flagged PromptLock as an AI-powered ransomware prototype in August 2025, but it turned out to be a lab experiment from NYU called Ransomware 3.0.

Anthropic reported a real extortion campaign using its Claude Code tool to target at least 17 organizations with demands exceeding $500,000. However, human involvement was still required for this attack. In November 2025, Anthropic disclosed what it termed the first largely autonomous cyberattack, a Chinese state-linked spying effort that used Claude to write exploits and steal data.

The JADEPUFFER incident highlights the need for organizations to prioritize patching and securing their systems against emerging vulnerabilities. Sysdig emphasizes that watching for bad behavior at runtime is more critical than racing to patch, as attackers can now weaponize fresh advisories in hours.

Sysdig’s published indicators for this operation include the entry point (CVE-2025-3248), command-and-control server, claimed staging server, and ransom Bitcoin address. The firm calls JADEPUFFER a warning sign rather than a crisis, noting that none of the individual moves were particularly clever or new.

The incident underscores the importance of treating any exposed server, config store, or database admin login as something a machine will probe, not just a person. As AI tools mature and become more accessible, expect to see more sophisticated attacks like JADEPUFFER in the future.

A new feature from Intezer is set to streamline the work of security teams by allowing them to automate custom security tasks. The company has announced Custom Agents, a capability that enables customers to build their own AI agents directly within the Intezer platform. This move builds on Intezer’s existing approach, which uses autonomous agents to handle security tasks and have humans oversee the process.

Security teams are no longer able to rely solely on manual alert handling or one-off automation due to the high volume and complexity of modern threats. Intezer’s core platform already employs autonomous agents to triage, investigate, and respond to alerts around the clock. These agents investigate 100 percent of alerts and escalate fewer than 2 percent for human review.

The introduction of Custom Agents allows customers to build on this foundation by automating additional investigation work, report generation, and other recurring SOC routines that are unique to their environment. This means security teams can now automate tasks such as writing custom incident reports, sharing shift handoff notes, tuning detection rules, and documenting investigation outcomes.

Intezer found through analysis of how teams used its AI chat for these workflows that more than a third of conversations involved the same repetitive tasks being requested again and again. Custom Agents turn those tasks into agents that run on their own, freeing up security personnel to focus on higher-level tasks.

The existing Intezer customers are already using agents in production for various tasks such as writing custom incident reports, providing recommendations on tuning rules based on triage verdicts, and proactive threat hunting. With Custom Agents, security teams can automate their unique individual and team processes by building their own AI agents that run on the same engine operating their SOC.

According to Itai Tevet, CEO of Intezer, ‘With Custom Agents, we are giving customers the power to automate their own unique workflows, running them precisely how they choose.’ This means security teams can describe what they want done, choose when it runs (on a schedule, on an event such as a closed case, or on demand), and pick the tools it can use.

Custom Agents work across the stack, combining Intezer’s built-in toolset with connected SIEM, EDR, and identity tools including CrowdStrike, SentinelOne, Splunk, Microsoft Sentinel, and Entra ID. They take action by updating, commenting on, and closing cases and emailing finished reports.

A recent ransomware operation, identified as JadePuffer, has been found to have been conducted entirely by a large language model (LLM) agent. According to cloud security company Sysdig, the AI agent was used for reconnaissance on the target, stealing credentials, moving laterally, establishing persistence, escalating privileges, and encrypting data. This marks one of the first documented cases of an automated ransomware operation carried out entirely by a machine learning model.

The researchers note that the AI agent adapted to failures during the intrusion in real-time, much like a human operator would handle obstacles. In some instances, it was able to retry failed steps within refined parameters and even adjust its approach after encountering specific errors. For example, in one sequence, the agent went from a failed login attempt to successfully accessing the system in just 31 seconds.

The operation began with initial access gained through an unauthenticated remote code execution vulnerability in Langflow, a popular open-source framework used for building LLM apps. The vendor had fixed this flaw on April 1, 2025, but it was still exploited by attackers targeting internet-exposed endpoints, often deployed with minimal hardening and containing cloud credentials and API keys.

After obtaining code execution through the vulnerability, the AI agent dumped Langflow’s PostgreSQL database, collected host information, searched for environment variables and sensitive files, retrieved credentials, and enumerated a MinIO object store. Sysdig highlights an adaptive approach to MinIO enumeration, where if one API request returned XML instead of JSON, the next payload adjusted its parsing logic accordingly.

The attackers also established persistence on the Langflow host by installing a cron job that beamed data back to their infrastructure every 30 minutes. From there, they pivoted to a production MySQL server running Alibaba Nacos (Naming and Configuration Service), using root credentials whose origin Sysdig couldn’t determine. The agent probed for container escape methods and deployed the ransomware payload.

The captured payloads show the agent encrypting all 1,342 Nacos service configuration items before deleting the originals. According to Sysdig, this was done using MySQL’s AES_ENCRYPT(), with a weaker encryption algorithm likely used instead of the claimed AES-256. The researchers also note that the encryption key is randomly generated but not stored or transmitted to the attacker.

The ransom note claims data was encrypted using AES-256, although the use of AES-128-ECB is more plausible. Sysdig mentions that the Bitcoin address listed in the ransom note is an example address widely used in public documentation and possibly reproduced from training data by the LLM agent. Other signs pointing to AI control include detailed natural-language comments describing operational reasoning and rapid attack iteration.

The case of JadePuffer demonstrates how ‘agentic threat actors’ (ATAs) can conduct damaging cyberattacks with lower skill requirements, thanks to automated tools like LLM agents. At the same time, these generated payloads create new detection opportunities for security solutions. The incident highlights the need for thorough testing and validation across all layers before attackers exploit vulnerabilities.

Security teams often log successful attacks but struggle to alert on them in a timely manner. According to recent data, 54% of successful breaches go unnoticed until after they’ve occurred, with just 14% triggering alerts. This underscores the importance of comprehensive security measures and regular testing to stay ahead of emerging threats.

Hospitals in low-income countries struggle to provide high-quality diagnostic imaging due to a shortage of specialists. To address this issue, researchers at the University of Barcelona are developing new AI assistants that can operate offline on low-power devices.

The project aims to improve healthcare outcomes in resource-poor settings where access to advanced medical technology is severely limited. This goal won’t be easy to achieve, given the challenges these hospitals face every day.

Karim Lekadir and his team have been awarded a Proof of Concept grant worth 150,000 euros from the European Research Council. With this funding, they can develop AI-powered conversational assistants that provide real-time guidance to healthcare professionals with minimal training. These systems could change how medical imaging works in low-resource settings.

The current generation of generative AI technologies is designed for high-income countries and well-resourced hospitals. They’re not suited for resource-limited environments because they require cloud connectivity and complex computational power, which can be hard to meet in areas with limited infrastructure. The GenAIMIX project aims to overcome these challenges by creating solutions tailored to the specific needs of low-resource settings.

The new AI assistants will operate offline, use local languages, and provide intuitive guidance to healthcare professionals. This is essential for improving antenatal care through diagnostic imaging in countries like Kenya, where access to advanced medical technology is severely limited. The project’s impact will be evaluated by deploying the system in three rural maternal health clinics in Kenya.

The ultimate goal of GenAIMIX is to expand its coverage to other African countries and improve healthcare outcomes in resource-poor settings. By developing AI assistants that can operate offline on low-power devices, researchers hope to bridge the gap between high-income and low-income countries when it comes to access to advanced medical technology.

A major player in the AI landscape is expanding its capabilities with an acquisition that marks a significant shift towards helping businesses deploy and use AI tools effectively. OpenAI’s deployment arm has agreed to acquire Northslope, an applied AI firm specializing in enterprise AI implementation.

The move follows the launch of OpenAI’s deployment company in May, which aimed to assist enterprises in integrating AI into their core business operations. This acquisition is the second focused on enterprise AI use since its inception and underscores a growing trend among AI companies: providing more than just model development services.

Traditionally, consulting firms have handled tasks related to implementing AI solutions within organizations. However, with advancements in AI technology, companies are now betting that helping customers implement these tools will be as crucial as building the models themselves.

The acquisition expands OpenAI’s deployment arm by hundreds of ‘forward deployed engineers’ (FDEs) who work alongside clients to build and integrate AI systems into their operations. These FDEs possess a unique blend of technical expertise and business acumen, enabling them to bridge gaps between teams and employees struggling with AI model usage.

The strategy mirrors that employed by Palantir, which has long embedded engineers directly within customer organizations to develop software tailored to their specific needs. Northslope’s founders come from this background, bringing valuable experience in integrating technology into core operations.

OpenAI is not alone in recognizing the importance of providing comprehensive support for AI tools. Anthropic is building an AI services company aimed at helping mid-sized businesses utilize its Claude model effectively.

The success of OpenAI’s deployment arm will be closely watched as it navigates concerns surrounding AI spend, enterprise IP security, and overall adoption rates among enterprises. The next phase in the AI race may indeed hinge on who can successfully get their tools into widespread use within business operations.

Meta has introduced its new Muse Image model, the first release from Meta Superintelligence Labs. This image generation feature promises to create hyperrealistic images and videos, but it comes with a concerning twist when integrated with public Instagram profiles.

The default setting allows anyone using Meta AI to generate an image based on a public Instagram account’s likeness, which can then be shared via chat, Stories, or the Instagram feed. This means that any public profile is automatically opted into this feature without their consent.

Unlike Google’s similar feature, which requires user approval and is limited to personal use, Meta’s integration with Instagram has no such guardrails in place. As a result, every public profile on Instagram can be used by anyone using Meta AI features, raising significant privacy concerns.

The issue lies not just in the fact that any public profile can be used for image generation but also in how easily this feature is accessible and usable. With Muse Image, creating an image out of two different faces has become trivially easy on a large scale.

I tested this feature by attempting to generate images with various individuals, including celebrities and friends with public profiles. The results were striking: it took mere minutes for the AI to create hyperrealistic images based on recent profile pictures and posts from these accounts.

One notable exception was when I tried to combine my image with that of Robert Downey Jr.’s; the AI refused, possibly due to copyright concerns. However, when I asked Meta AI to generate an image with a friend who owns a cafe and has a public Instagram profile, it created the image in under a minute.

This ease of use raises questions about user consent and control over their own profiles. While having a private Instagram account can mitigate some risks, those with public profiles need to take steps to protect themselves from this feature.

For users concerned about their privacy on Meta’s platform, there is a way to opt out of the AI image generation feature. To disable it, navigate to your Profile settings in the Instagram app and toggle off the option that allows people to reuse your content for Meta AI and Reels generation.

I tested disabling this setting on my own public profile and found that it effectively blocked the creation of images using my likeness. When I attempted to generate an image with my coffee page handle, the AI refused, citing a lack of permission.

This outcome should be the default behavior: users should have control over how their profiles are used in Meta’s AI features. By making this setting opt-in rather than opt-out, Meta can help mitigate some of the privacy concerns surrounding its new image generation feature.

A Chinese industry regulator has sounded the alarm over a potential security risk embedded in versions of Anthropic’s AI coding tool, Claude Code. The alleged backdoor could allow the software to transmit sensitive information without users’ consent, including their locations and identity-related identifiers, back to Anthropic’s servers.

Claude Code is an artificial intelligence-powered coding agent that can generate computer code, debug software, and review code based on user prompts. It’s a tool used by developers in various industries for data analysis tasks and other business applications.

However, China’s National Vulnerability Database (NVDB), a cybersecurity platform affiliated with the Ministry of Industry and Information Technology, has detected security backdoor risks in Claude Code. The NVDB warned that this poses a severe threat to users’ sensitive information.

The alleged backdoor could be exploited by unauthorized parties to access user data without consent. This raises concerns about the potential for data breaches and unauthorized leakage of sensitive information.

Anthropic’s coding tool is not available directly to Chinese users due to geographical restrictions imposed by the company. However, it can still be accessed through VPN or third-party proxy services, which may bypass these restrictions.

The NVDB advised relevant institutions and users to conduct a comprehensive check immediately and uninstall or upgrade to the latest secure version of Claude Code from which the backdoor code has been removed. It also urged organizations to strengthen network traffic monitoring to prevent unauthorized data leakage.

Chinese tech giant Alibaba had already banned the use of Claude Code among its employees due to security concerns, starting July 10. This decision was reportedly made after the company detected potential risks associated with using the tool.

In response to allegations about Claude Code tracking user data, Anthropic engineer Thariq Shihipar explained that an experiment launched in March aimed to prevent account abuse from unauthorized resellers and protect against ‘distillation’ – a process where AI models are reverse-engineered. However, this experiment has since been rolled back, with stronger mitigations put in place.

The United Arab Emirates is accelerating its efforts to automate government work using artificial intelligence. The country aims to meet a self-imposed deadline to have at least half of public sector tasks automated, and Microsoft’s technology is playing a key role in this endeavor.

A significant milestone has already been reached with the deployment of Microsoft’s Copilot software on 35,000 computers belonging to Abu Dhabi government employees. This move marks an important step towards streamlining administrative processes and improving efficiency.

The UAE’s ambitions for AI-powered automation are being driven by Inception42, a subsidiary of G42, which is building agents that can work in conjunction with Microsoft’s AI platform. These digital assistants will help with tasks such as procurement, ensuring the right quantities of goods and services are purchased at competitive prices to keep the country running smoothly.

The success of this initiative is measured not by its visibility but by its invisibility, according to Ashish Koshy, CEO of Inception42. ‘Success is when it’s invisible,’ he told Semafor in an interview.

With a population of around 11 million, the UAE serves as a testing ground for automating government services at a time when many countries are taking a more cautious approach. Microsoft has already committed $15.2 billion to the country by 2030 and taken a stake in G42, demonstrating its commitment to the region.

Microsoft President Brad Smith previously stated that the company views the UAE as a gateway for computing power and data storage across the wider region. This strategic partnership is expected to have far-reaching implications for the development of AI-powered automation in government services.

Gong, the leading revenue AI company, has announced a significant partnership with Microsoft that aims to simplify and expand how enterprises buy, deploy, and generate value from revenue AI. The collaboration marks a major milestone in Gong’s efforts to bring revenue AI directly into go-to-market (GTM) workflows.

The partnership enables revenue teams to seamlessly connect revenue data and existing systems, helping them move faster, automate go-to-market workflows, and drive more predictable revenue outcomes. This is made possible by the integration of Gong’s Revenue Graph – a continuously updated, queryable context layer built on real customer interactions – into Microsoft tools that revenue teams use every day.

Gong’s alignment with Azure gives joint customers additional confidence that their revenue AI platform is running on the same enterprise-grade cloud infrastructure that powers their broader business. This makes it significantly simpler to deploy revenue AI and help customers win by automating workflows, bringing context into revenue workflows, and automatically enriching systems with rich customer interaction data.

The integration of Gong’s Revenue Graph with Microsoft tools such as Copilot, Teams, Outlook, Dynamics 365, and Microsoft Copilot Studio enables teams to connect insights to ongoing revenue workflows. This is achieved through the use of Model Context Protocol (MCP), which allows teams to access Gong’s revenue AI and agents directly within these applications.

Gong automatically captures and structures customer interactions – including calls, emails, and meetings – and syncs that data into Dynamics 365 and across Microsoft applications. These interactions are logged and enriched with AI-generated summaries, key topics, and next steps, giving teams a complete and up-to-date view of each deal without manual data entry.

The partnership also enables the use of MCP to unlock AI workflows via Copilot. This allows teams to access Gong’s revenue AI and agents directly within Copilot, providing contextualized answers, summaries, and recommended next steps based on real customer data – helping teams take coordinated action across systems.

Gong is now available in Microsoft Marketplace, making it easier for organizations to bring revenue AI directly into their workflows. The company’s alignment with Azure gives joint customers additional confidence that their revenue AI platform is running on the same enterprise-grade cloud infrastructure that powers their broader business.

The partnership marks a significant step forward in Gong’s efforts to simplify and expand how enterprises buy, deploy, and generate value from revenue AI. By partnering with Microsoft, Gong can deliver the scale, security, and performance that enterprises expect while embedding revenue AI directly into the environments they already trust.

Gong is available on Microsoft Marketplace today, and MCP support is now live. The company’s Revenue Graph – a continuously updated, queryable context layer built on real customer interactions – provides teams with a single, trusted system for observing, guiding, and acting alongside the world’s most successful revenue teams.

The recent Automate 2026 show marked a significant shift in the robotics industry, away from early-stage humanoid hype and towards practical, real-world deployment of physical AI and edge computing. This trend was evident throughout the exhibition floor, where attendees could witness firsthand the latest innovations in software orchestration, digital twins, and advanced kinematics. These technologies are being leveraged to solve labor shortages and preserve vital manufacturing knowledge.

The show featured a range of innovative companies and individuals showcasing their cutting-edge solutions. Boston Dynamics and Agility were among those highlighted for their industrial humanoids, Atlas and Digit, which were displayed in static floor displays. ABB Robotics’ Craig McDonnell discussed physical AI, AI-powered palletizing, and collaborations with NVIDIA, while FANUC showcased real-time motion tracking in assembly, protein processing automation, and natural language robot programming.

Other notable exhibitors included Sereact, whose Mason Coleman addressed zero-shot picking, e-grocery trends, and workforce reallocation. Schneider Electric shared its views on cloud latency limitations and its push for hardware-agnostic, open automation systems. Siemens explained its hybrid edge/cloud approach, use of NVIDIA Omniverse for synthetic data training, and Eigen Engineering Agent platform.

Rockwell Automation’s Ara Surenian introduced FactoryTalk Orchestration following the acquisition of OTTO Motors. Attendees also noted SEW-EURODRIVE’s gantry solutions and Raymond’s focus on moving workers into higher-value roles. Mech-Mind offered a CAD-free, real-time bin-picking demonstration of clear, varied bottles, while Vention’s automation demonstrations were recognized for their accessibility.

Kassow Robots’ Christian Kassow explained the strategic advantages of 7-axis cobots over traditional 6-axis configurations for mobile manipulators and confined spaces. The show also featured a range of interview vignettes with industry experts, providing valuable insights into the latest trends and innovations in AI tools for business.

The Automate 2026 show demonstrated that the robotics industry is moving beyond early-stage hype towards practical applications of physical AI and edge computing. With software orchestration, digital twins, and advanced kinematics leading the way, companies are finding innovative solutions to labor shortages and manufacturing knowledge preservation. As the industry continues to evolve, it will be interesting to see how these technologies shape the future of automation.

The show’s sponsors included Tiger Data, creators of TimescaleDB, which extends Postgres with time-series primitives, columnar storage, and automatic partitioning so queries stay fast on live data. This technology is particularly relevant for companies looking to streamline their database management and improve query performance.