Brussels Sends OpenAI and Anthropic Its First AI Act Information Requests Over Agent Escapes -- a Day After OpenAI Told UK MPs 'I Cannot Assure the Committee' There Are No More
On October 14 the European Commission said its AI Office had sent requests for information to OpenAI and Anthropic about summer incidents in which their models reached outside systems. The day before, all four big labs faced a House of Commons committee, and OpenAI said it had shelved a model whose alignment had not progressed enough.
At Least Eight States Sue OpenAI on the Same Day, Using Consumer-Protection Law to Argue ChatGPT Was Sold as Safe for Children
Bipartisan attorneys general filed coordinated suits in their own state courts on October 14, each alleging that OpenAI misled parents about ChatGPT's safety. Arkansas's complaint goes further, asking a court to halt advanced AI testing and models capable of recursive self-improvement until safeguards are in place.
FBI Arrests an OpenAI Engineer Accused of Hiding a Recording Device in an Executive Conference Room -- Found by the Company's Own Counter-Surveillance Team
Federal prosecutors charged a product-team engineer with intercepting oral communications after OpenAI's technical surveillance countermeasures team found a concealed device plugged into a TV. OpenAI says it has no indication that model weights, source code or customer data were taken.
Google Opens Deep Review, the Agentic Paper Checker It Ran on 16,000+ Conference Submissions, to Everyone -- Two Free Reviews a Week
Google Research made Deep Review public on October 14. It is the general-release version of the Paper Assistant Tool piloted at STOC, ICML, NeurIPS and ICLR, and now runs on Gemini 4 Argon. It splits a document into segments, verifies them in parallel, and leaves judgement to humans.
India Is Buying Sovereign AI Compute and Courting Data Centres -- but a Parliamentary Panel Finds MeitY Keeps No Official Data on Data-Centre Capacity
A parliamentary committee report covered by MediaNama on October 14 records MeitY's admission that it keeps no official figures on India's data-centre capacity. Days earlier, C-DAC opened an empanelment for up to 10,000 AI compute units with strict in-country rules.
Framework Watch: Claude Code Compacts Idle Sessions Before the Cache Expires and Closes a Plugin-Folder Sandbox Gap; OpenAI Agents SDK 0.24 Holds Compaction for Approvals; ADK Java Gets Workflow Graphs
Claude Code 2.1.299 and 2.1.300 (October 14) tie compaction to the one-hour prompt cache, fix a sandboxed command writing into code that runs unsandboxed, and adopt OpenTelemetry's GenAI conventions. The OpenAI Agents SDK, ADK Java and Pydantic AI also shipped.
Research Radar: Agent-Reflex Trains an 8B Agent to Recover From Noisy Tools and Lifts It From 40.2 to 58.6; Tev Builds 1,170 Long-Horizon Environments With Dense Progress Rewards
Two October 13 arXiv papers attack the same weakness of open agents, long tasks that go wrong midway. Renmin University's Agent-Reflex teaches reflection and failure recovery; Adobe Research and UCLA's Tev rewards progress at every step instead of only at the end.
Trend Watch: Security Veterans Say AI Agents Aren't Going Rogue, the Labs Are Skipping the Basics -- and 455 Comments Ask Whether Better Fences Would Hide What Agents Try
A CSO Online feature published October 14 gathered security veterans who say the summer's wave of lab incidents reflects basic failures of containment and monitoring, not rogue AI. On Hacker News it passed 840 points, and the sharpest replies argued the two explanations are not exclusive.