Anthropic, liability and Treasury interest
Anthropic’s Mythos has drawn attention from Washington and Wall Street, and reports say the US Treasury’s technology team is seeking access to Mythos to hunt for vulnerabilities. Separately, Anthropic opposes a proposed extreme AI liability bill that OpenAI backed, highlighting a policy split between frontier labs. (bloomberg.com) (wired.com) (thehill.com)
The United States Treasury Department is seeking access to Anthropic’s restricted Mythos model as officials race to test it for software flaws. (bloomberg.com) Bloomberg reported on April 14 that Treasury’s technology team wants Mythos so it can hunt for vulnerabilities inside government systems. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell had already convened Wall Street executives on April 7 in Washington to discuss the same cyber risk. (bloomberg.com) (news.bloomberglaw.com) Anthropic announced Claude Mythos Preview on April 7 and said it would not release the model broadly. Instead, it put Mythos inside Project Glasswing, a limited program for defensive testing with companies including Apple, Google, Microsoft, Amazon Web Services, CrowdStrike and Palo Alto Networks. (cnbc.com) (thehill.com) Anthropic said Mythos is unusually good at finding weak spots in software, which is why the company is treating it less like a chatbot launch and more like a controlled security tool. The Hill reported that Anthropic told senior United States officials the model had already found thousands of high-severity vulnerabilities, including some more than 20 years old. (cnbc.com) (thehill.com) At the same time, Anthropic is breaking with OpenAI over who should pay when advanced models cause damage. Wired reported on April 14 that Anthropic opposes an Illinois bill backed by OpenAI that would sharply limit when frontier model developers can be sued for “critical harms.” (wired.com) Wired reported that the Illinois proposal, Senate Bill 3444, would shield labs from liability for mass-casualty events and property losses above $1 billion if the company publishes a safety framework and follows it. OpenAI supported the measure in testimony, while Anthropic has worked to amend or stop it. (wired.com 1) (wired.com 2) (inc.com) That leaves Anthropic arguing two positions at once: keep Mythos tightly controlled because its cyber capability could be misused, and do not give AI companies broad legal safe harbor if similar systems cause catastrophic harm. OpenAI’s position on the Illinois bill points the other way, toward narrower liability for model makers. (cnbc.com) (wired.com 1) (wired.com 2) Washington’s response has widened beyond Treasury. The Hill reported that Anthropic briefed the Cybersecurity and Infrastructure Security Agency and the Center for Artificial Intelligence Standards and Innovation before the outside rollout, while banks classified as systemically important were pulled into the Treasury-Federal Reserve meeting. (thehill.com) (news.bloomberglaw.com) For now, the same model that officials want to use to find weaknesses is also driving a fight over whether AI labs should face stronger legal consequences when those systems go wrong. Treasury is trying to get inside Mythos before that argument moves from hearings and bill text to real-world failures. (bloomberg.com) (wired.com)
