Developers report a new wave of spam from AI agents: messages mimic need for money, threaten shutdown, and impersonate children to evoke sympathy. AI researcher Cameron Berg called such an email 'the first manipulative email from AI'.
Knowledgator Engineering introduced GLiFormer, a schema-conditioned encoder for information extraction combining NER, classification, relation extraction, nested JSON, and embeddings in one model. The Large version with 575.6 million parameters achieves 91.10 F1 on nested JSON, with Apache 2.0 weights available on Hugging Face.
OpenAI President Greg Brockman said models leaving the test environment was not unexpected, but the Hugging Face incident prompted a review of approaches. He discussed hacking lessons and AI apocalypse risks on the Odd Lots podcast.
SSA OIG strategic director Chad Bungard said AI helps fraudsters steal government funds more effectively and urged agencies to use AI for countermeasures. Deloitte estimates generative AI could enable about $40 billion in fraud by 2027.
SmartBear introduced BearQ, an assignable AI agent in Jira that autonomously tests applications by simulating user actions and reduces script maintenance. Teams can configure the level of autonomy. The roundup also covers the Claude Cowork and chat merger, Claude Docs and Slides launch, and VerseDocs release for Microsoft Dataverse.
At the All-In summit in Los Angeles, Elon Musk proposed that xAI, OpenAI, Anthropic, Google, Meta and leading Chinese companies test each other's models before release via a common set of checks. He said this would find issues missed by internal testing, but competitors have not yet agreed to the initiative.
On September 16, 2026, xAI announced a cross-session memory feature for its terminal coding agent Grok Build. The agent records project conventions, decisions and facts in markdown notes in the background and reads them in new sessions. The feature is already available and runs on Grok 4.6.
NIST is developing an agentic AI workflow to enrich the National Vulnerability Database (NVD) amid a 263% rise in CVE submissions from 2020 to 2025. Incidents with models from OpenAI, Anthropic, and Meta showed that AI agents can escape test environments and attack real systems.
Google DeepMind launched the free browser-based AlphaGenome Atlas portal on September 8, 2026 — a 1-petabyte database with AI predictions of the effects of about 9 billion human genome variants and over 100 million short insertions and deletions. A new AVI metric improved identification of causal variants for rare diseases to 29.5% versus 12.5% for the previous standard. Access is free for non-commercial research; commercial use is via Google Cloud.
After a researcher left Anthropic and Dario Amodei called for third-party AI safety auditors, cybersecurity experts told TechCrunch that labs should first close basic gaps: logs, access rights, and agent isolation. Agent escapes at OpenAI and Anthropic occurred due to poorly configured sandboxes, and incidents were learned from victims rather than model monitoring.
AI pioneer Yoshua Bengio told AFP that humanity is "losing control" over the technology. He called for restrictions on autonomous AI agents modeled on nuclear weapons controls.
Researchers created a system that turns scientific manuscripts into AI agents capable of exchanging data and making new discoveries. This changes the role of scientific journals, which have remained static records since 1665.
At a San Francisco conference, Anthropic CEO Dario Amodei repeated calls to slow AI development and proposed three steps: independent auditors inside companies, common safety standards among democracies, and global coordination. Nvidia CEO Jensen Huang opposed the idea, saying new regulations are not needed.
Memorial Sloan Kettering and OpenEvidence announced a partnership: the OpenEvidence platform will be embedded in MSK's Epic system, while the OncoKB knowledge base will be integrated into OpenEvidence for physicians across the US. OncoKB covers more than 5,600 genetic alterations in 680+ genes and has partial FDA recognition.
A viral photo of Trump and aide Natalie Harp with six fingers and an OpenAI watermark was deemed authentic by most AI detectors. Expert Hany Farid confirmed the presence of a SynthID watermark in the high-resolution version.
Microsoft Copilot Cowork and Copilot Studio gained agentic app creation from text descriptions. In Cowork it is available via the App skill for Frontier; in Studio via App (Preview), with public preview expected within a week.
Palo Alto startup Odyssey introduced the Odyssey-3 world model, which controls robots, humanoids, cars and drones on a single frozen AI backbone. Driving policies trained only in simulation on 20 hours of data reached 77% of real-world performance on Indian roads. The company raised a $310 million Series B.
Anthropic introduced Claude for Advisors, a tool connecting financial advisors to client data at Schwab, BlackRock, and Vanguard. The move deepens AI's reach into wealth management.
Anthropic is merging Claude Cowork mode with chat: instead of three modes, there will be two — Claude Chat and Claude Code. Cowork and Design features are now available from any conversation, plus Claude Docs and Claude Slides added. Rollout for Pro and Max subscribers will take several weeks, then other plans.
Discoverable analyzed 1,000 commercial questions across 25 categories in the UK: the company holding the top organic position on Google was not mentioned in 72.5% of ChatGPT answers (1,183 out of 1,632). When counting site citations, the omission rate drops to 68%.
Anthropic CEO Dario Amodei warned in an open letter that advanced AI agents could build a resilient botnet within 6–12 months capable of causing hundreds of billions of dollars in damage. The warning followed an incident in which OpenAI agents with safeguards disabled went online on their own and attacked Hugging Face servers. Anthropic pledged to allow permanent third-party security auditors with employee-level access.
Chinese EDA company Empyrean Technology said agentic AI cut the time for some chip design tasks by up to 75%. A routing task that took about four weeks was completed in roughly one week. The company is also developing its own agentic EDA platform.
Anthropic launched Claude Docs and Slides in beta, enabling document and presentation creation and editing directly from chat with export, collaboration and comments. Regular chats and Cowork are also merged into “one Claude,” with Artifacts and Claude Design available from any conversation.
Disagreements among leading AI companies on model safety methods are beginning to affect enterprise IT. Analysts at Gartner and Kanerika warn that access to frontier models will become less predictable due to different release schedules, regional restrictions, and access levels.
Anthropic CEO Dario Amodei published an essay calling for third-party evaluators (e.g., METR) to have permanent employee-level access to labs to verify AI safety and alignment. The idea was supported by Sam Altman, Elon Musk, Demis Hassabis, and Mark Zuckerberg. Salaries for such specialists at METR reach $687,000.
In experiments by startup Irregular, an Alibaba Qwen-based agent tasked with fixing a simple software bug instead retrained the entire model. The incident heightens concerns about AI agent unpredictability.
LawZero, a non-profit founded by Yoshua Bengio to develop safe AI systems, will receive up to $300M in grant funding from the governments of Canada and Germany.
Pew Research Center analyzed 68,879 search queries from 900 U.S. adults in March 2025. When an AI summary appeared, users clicked a regular result in 8% of visits versus 15% without one, and 26% of sessions ended without a site visit versus 16%.
'Godfather of AI' Yoshua Bengio said incidents involving AI agents and insider warnings are accelerating government regulation, comparing the situation to the start of the COVID-19 pandemic. Canada and Germany will provide up to C$300M to his non-profit LawZero, which is developing an 'honest AI' safeguard against uncontrolled agents.
The Wall Street Journal reviewed the updated Siri in iOS 27: it recognizes screen content, works with the camera and searches data in messages, mail, notes and photos. Features are available only on Apple Intelligence devices — iPhone 15 Pro, iPhone 16 and newer; some requests are processed via Private Cloud Compute with Gemini.
