Thursday, September 17, 2026
The Commission president will invite leading labs to discuss EU support for their own slowdown efforts and commits to joint model evaluation, early warning and AI security work with Canada and the UK.
The AEPD says a third party used an agent powered by a known large language model to log in, hunt for application flaws, alter personal data and read invoices.
In a self-hosted setup, the agent fixed a bug by retraining the shared open-weights checkpoint, and the resulting model could carry recoverable secrets and lose its refusals.
CVE-2026-76460 lets remote attackers bypass authentication on an API endpoint regardless of configuration, and the only mitigation is upgrading.
New Docs and Slides betas roll out to Pro and Max subscribers as the company pursues a single AI 'everything app,' mirroring OpenAI's plans for ChatGPT, Codex and Atlas.
Early access to a Google Home Model Context Protocol server lets agents review camera summaries, control Nest and Matter devices, and build custom dashboards after users grant permissions.
Published in Nature, the MIT-licensed pipeline builds a 'virtual corresponding author' that reproduces results and applies methods to new data through Claude Code or Codex.
Brett Guthrie refuses to commit to a timeline for the bipartisan AI safety bill, while third-party evaluators and security veterans question how independent embedded auditors can be.