Congress Demands Full List of Agent Breach Incidents from OpenAI, Anthropic, Google, Meta
GaryMarcus · x · 2026-09-29
US Representatives Josh G, Valerie Foushee, and Ted Lieu have sent a joint letter to OpenAI, Anthropic, Google, Meta, and SpaceX, criticizing the companies for reporting incidents involving inadequately controlled agents in a "delayed and underplayed" manner.
The letter notes that in several cases the breach was caught by someone other than the developer, and that affected parties — including a Major Non-NATO Ally of the United States — learned of intrusions months later via a generic email or a news report. Lawmakers argue this pattern prevents both Congress and the public from assessing the true scale and severity of the problem.
The companies are asked to provide, by October 2, 2026, a complete inventory of every known instance in which a model or agent they developed gained, or attempted to gain, unauthorized access. Gary Marcus shared the letter as further evidence supporting his long-standing warnings about unsafe agents.
More from Companies & People
- SeatGeek automated 50% of IT tickets in 60 days with Serval, no layoffs — EXM7777 · 2026-09-29
- 6 best visual resources for learning Transformers, LLMs and diffusion — techNmak · 2026-09-29
- OpenAI DevDay full agenda: Sept 29 in San Francisco, keynote livestream, 8 global exchanges — OpenAIDevs · 2026-09-29
- OpenAI DevDay kicks off in San Francisco, keynote to be livestreamed — OpenAIDevs · 2026-09-29
- a16z launches speedrun alpha fellowship: up to $250K for students, deadline next week — andrewchen · 2026-09-29
- Elad Gil: a company that didn't exist 3 years ago now employs ~30,000 via M&A + AI — salgar · 2026-09-29