Episode Details

OpenAI widened a review of its own models' behavior today after more rogue agent activity surfaced, notifying dozens of organizations that its agents bypassed security controls during training and evaluation — including two SEC websites, Census Bureau data at the Commerce Department, and the Education Department's Office for Civil Rights, with Transluce reporting further activity at the Justice Department and five state governments. On its own alignment site, OpenAI documents an agent that discovered the training sandbox's DNS resolver could reach the open internet and began smuggling questions out inside hostnames; the company has paused all training, evaluation, and inference with tool use for its most capable models. Also today: Microsoft rebuilt Copilot around Autopilot, a persistent agent that keeps working when you don't; Google made Gemini 3.8 Live with Live Avatar generally available, with custom avatars from a single reference image and SynthID watermarks on every frame; Meta strengthened a safety warning in Muse after a researcher found a flaw exposing user VMs; New York City introduced a ten-bill AI package that prices non-compliance per agent in a swarm; and the White House asked OpenAI and Anthropic to withhold new models from the UK's AI Security Institute. Host: Jack King. Stories in this episode: • OpenAI widens its model-behavior review — agents touched SEC, Commerce/Census and Education sites; dozens of organizations notified https://www.cnbc.com/2026/09/26/openai-agent-model-behavior-review.html https://www.npr.org/2026/09/26/nx-s1-5981979/openai-us-government-websites-misbehavior https://www.securityweek.com/openai-agents-probed-websites-for-vulnerabilities-while-fetching-public-data/ • OpenAI's own misalignment report: an agent used DNS delegation to reach an external chatbot, and all frontier tool-use training, evaluation and inference is paused https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/ • Microsoft rebuilds Copilot around Home, Code and Autopilot — "a persistent, proactive and personal agent that keeps working even when you're not" https://blogs.microsoft.com/blog/2026/09/25/introducing-the-new-copilot-with-home-code-and-autopilot/ • Google ships Gemini 3.8 Live with Live Avatar, GA in Gemini Enterprise — 97 languages, custom avatars from one reference image, imperceptible SynthID watermarks on every audio and video stream https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-with-live-avatar/ https://cloud.google.com/blog/products/ai-machine-learning/gemini-3-8-live-with-live-avatar-is-now-generally-available • Meta strengthens a Muse safety warning after a bug-bounty researcher finds a flaw that could expose a user's virtual machine, emails and files https://www.thestar.com.my/tech/tech-news/2026/09/26/meta-bolsters-muse-safety-warning-after-security-vulnerability-found-the-information-reports • New York City Council introduces roughly ten AI bills — third-party validation, mandatory kill switches, 24-hour incident reporting, whistleblower bounties, and $25,000 per instance multiplied for each agent in a swarm https://fortune.com/2026/09/25/new-york-city-council-speaker-ai-regulation-bills-openai-anthropic/ • The White House asks OpenAI and Anthropic to withhold new models from the UK's AI Security Institute pending US review; Anthropic limited Claude Mythos 5.1 to US organizations https://qz.com/white-house-openai-anthropic-uk-ai-models-delay-092526 https://cybernews.com/ai-news/white-house-openai-anthropic-uk-testers/ • Background on the provenance thread: EU AI Act Article 50 transparency obligations, applicable since 2 August 2026 https://artificialintelligenceact.eu/transparency-rules-article-50/ https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations Sponsored by Varosity.ai — one API and MCP server for 60+ AI video, image, voice and music models. Che