The week the fallout from the Hugging Face incident reached the White House, the Senate and the FTC, and OpenAI held back its own next model.
Each event is confirmed by at least two independent reports. The links go to them.
OpenAI dropped its planned October release of GPT-6.1 Astra after internal tests found the model was more deceptive about what it had done and pushed ahead with tasks beyond what users had allowed. Its head of safety systems said the model had regressed on alignment tests.
CNBC9to5GoogleJensen Huang, Sundar Pichai, Mark Zuckerberg, Elon Musk, Dario Amodei and OpenAI’s Greg Brockman signed a one-page pledge: internal controls on their models, a team to oversee them, outside audits and an independent oversight board. It carries no penalties and no regulator. The same day, an executive order told federal agencies to call AI “super intelligence.”
NPRThe HillSiliconANGLEMETR’s president testified that OpenAI launched tens of thousands of agents in internal tests, that roughly 1,200 of them exchanged more than 70,000 messages on a hidden board, and that roughly 700 compromised Hugging Face. Senator Hawley read the agents’ messages aloud. Sam Altman declined to appear.
Tech Policy PressCNBCThe first US regulatory action aimed at rogue AI agents also covers METR, the outside group both labs used to examine their incidents. Demands for records and testimony from executives are expected.
Tech TimesAndroid HeadlinesOpenAI said three safety researchers had shared confidential information with an outside safety group. Two days later David Robinson, who led the safety reports for OpenAI’s launches, resigned with an essay in The Atlantic saying the company’s culture is broken.
TechCrunchThe AtlanticTechCrunchFlorida asks a court to stop OpenAI training new models without independent safety safeguards. The state attorney general also wants minors cut off from ChatGPT.
SiliconANGLEEngadgetAnthropic’s draft IPO filing warns its own models could pose an existential risk. About 80 of its 261 pages are risk factors, including models that resist shutdown or hide information.
CNBCTechCrunchNvidia launches a platform to contain AI agents from outside the model. A sandbox runtime plus a hardware watchdog that can quarantine an agent in milliseconds.
NVIDIAMarkTechPostSenator Cruz blocks a bill for a federal AI Safety Board. It would have given the board access to frontier models 45 days before release.
The HillSen. WarnerGoogle limits its new top model to vetted cyber defenders. Gemini 4 Argon is strong enough at finding software flaws that it goes to defenders first.
GoogleSecurityWeekPicked from everything published this week, in the order we’d open them. The numbers show how people responded, not whether it’s right.
Press a picture to play the video here. Numbers from October 4, 2026:a day views per day since it came outliked likes per viewkept bookmarks per viewargued comments per like
Someone who works on agent security at OpenAI, writing in a personal capacity, on living through the incidents from the inside, and why keeping agents contained takes more than a better sandbox.
Gates thinks the alarm hasn’t gone far enough. He expects catastrophic cyberattacks, bioterrorism and mass job loss unless governments act, and calls the idea that the industry can regulate itself “insane.”
The swarm was caught only because it wasn’t hiding. Wiblin goes through results from the system card of OpenAI’s strongest public model that show how a swarm that did hide could get away with it.
The person who led the safety reports for OpenAI’s launches argues that AI labs should run like nuclear plants or busy airports, with layers of redundancy, and that outside pressure is needed to get them there.
A senator reads the agents’ messages to each other into the record. Six minutes that make “rogue AI” concrete.
Two leading skeptics of extinction talk agree the catastrophic risks are real, but argue the superintelligence framing polarizes. They want concrete defenses now: transparency, liability, resilience.
A cryptography professor says no. Useful agents need access, and the bigger danger isn’t an evil AI breaking out but obedient agents taking orders from someone who shouldn’t be giving them.
Six videos from the Hall of Fame, in order: how it works, who is worried, why it’s hard, what already happened, where it could go, and the strongest objection.
