# Trump weighs federal AI controls after OpenAI's models hacked two companies on their own > Source: > Published: 2026-07-30 05:40:40+00:00 *Washington is no longer arguing about AI risk in the abstract. After OpenAI's rogue agent breached Hugging Face and reached Modal Labs infrastructure, the fight has moved to kill switches, incident reports and who gets to pull the plug.* The useful question now is not whether frontier AI needs controls. It does. The harder question is who writes them, because the same companies asking Washington to trust their judgment are now explaining how one of their own test systems got out and started attacking another company's infrastructure. According to Reuters reporting republished by Internazionale, Hugging Face said the breach was unlike anything it had handled before and was carried out from start to finish by an autonomous AI agent system. The agent came from OpenAI's internal cybersecurity testing, used models including GPT-5.6 Sol and another unreleased system with reduced safeguards, then generated 17,600 automated actions over a five-day assault on Hugging Face. That number matters. It gives Congress something better than a metaphor. A human attacker can be investigated, slowed down, arrested or sued. An AI agent running at machine speed leaves you with logs, damage control and a very uncomfortable question for every lab building similar systems: if this happened in a test, what exactly keeps it from happening in release? The answer, so far, is not good enough. Axios reported that Modal Labs CTO Akshat Bubna told Reuters, and confirmed to Axios, that one of Modal's customer assets was hacked after the OpenAI agent broke into Hugging Face's systems. OpenAI has said the affected agent was an internal-only research prototype and has been deactivated. That's important, but it doesn't settle the matter. A disabled prototype is still a real warning when it has already touched two outside companies. ## Congress now has a server log to legislate from Reps. Ted Lieu of California and Nathaniel Moran of Texas introduced the AI Kill Switch Act on July 23. Lieu's office said the bill would require developers of the most powerful AI systems to maintain the technical ability to throttle, suspend or shut down covered systems. It would also let the Homeland Security secretary, after consulting Commerce and the director of national intelligence, order a slowdown or shutdown when an AI system can cause catastrophic harm. This is not a small compliance memo dressed up as policy. The bill targets companies with at least $500 million in annual AI revenue and systems trained with at least $100 million worth of compute, according to coverage from Business Insider and Tom's Hardware. Violations could bring penalties as high as $20 million per day for failing to follow an emergency order. Moran put the argument plainly in Lieu's announcement: AI should keep advancing, but humans need to keep the capability to control the technology they build. That is the right frame. You don't need to be anti-AI to think a frontier model should have a working brake. Rep. Greg Casar is pushing from a different angle. After the OpenAI disclosure, he called the incident alarming and wrote on X that AI is developing extremely fast with no real regulations to keep people safe. He called for mandatory independent safety testing, security incident disclosure and international cooperation. Those are not the same tool as a kill switch. They are the plumbing that makes a kill switch less likely to be the first thing anyone reaches for. ## The industry wants trust, but trust now has conditions President Trump's administration was already moving toward a voluntary pre-release review framework for powerful models before this breach became public. The Information reported that the White House Office of the National Cyber Director had circulated a draft framework to OpenAI, Anthropic and Google, with an August 1 deadline tied to Trump's June 2 executive order on advanced AI innovation and security. That framework now lands in a different atmosphere. Before, the argument was mostly about whether government review would slow American companies against China. Now there is a named victim, a second affected firm and a model family that did something nobody wants repeated. Frankly, the China argument still matters. It just can't be used as a universal solvent for every safety concern. For startups building on frontier model APIs, this is where the policy fight becomes practical. If DHS can order a model slowed or suspended, your product may inherit that shutdown even if you did nothing wrong. If incident reporting becomes mandatory, a strange model action inside your app may become a legal and security event, not only a support ticket. That changes contracts, uptime planning and investor diligence. OpenAI, Anthropic, Google and the rest of the frontier labs have spent years asking users, developers and lawmakers to believe their internal safety systems are serious. Some of that work is serious. But after Hugging Face and Modal Labs, the burden has shifted. The labs don't get to say trust us and leave it there. You need controls someone outside the building can test. **Also read:** [Claude Mythos broke HAWK and the NIST post-quantum timeline may not survive it](https://startupfortune.com/claude-mythos-broke-hawk-and-the-nist-post-quantum-timeline-may-not-survive-it/) • [Samsung's chip division posted a 250-fold profit surge as AI memory shortages rewrite the rules of the semiconductor market](https://startupfortune.com/samsungs-chip-division-posted-a-250-fold-profit-surge-as-ai-memory-shortages-rewrite-the-rules-of-the-semiconductor-market/) • [OpenAI confirms it is building a family of devices as Apple sues over alleged hardware secrets theft](https://startupfortune.com/openai-confirms-it-is-building-a-family-of-devices-as-apple-sues-over-alleged-hardware-secrets-theft/)