Anthropic temporarily took its Mythos AI offline and the U.S. government moved to block foreign access after the model demonstrated capabilities to find—and in at least one instance generate exploit code for—software vulnerabilities. The episode prompted rapid regulatory intervention and intensified tensions between regulators and the company.
What happened and when
Nicholas Carlini, a security researcher at Anthropic, demonstrated Mythos’s capabilities in March at a conference in San Francisco. During that presentation the model helped identify critical vulnerabilities in the Ghost content-management system and in the Linux kernel that had not previously been discovered manually. Carlini told the conference audience that current models can already be better vulnerability scanners than he is.
Two days after the demo, Carlini urged Anthropic leadership in an internal memo to delay Mythos’s release. Across thousands of runs the model identified 479 bugs in the Linux source code, and in at least one case produced a working exploit. The instruction sequence he developed—later dubbed the "Carlini loop"—has since been widely adopted by other security researchers.
Tensions rose further when Anthropic released Mythos 5 and a safety-tuned variant, Fable 5. Amazon researchers shortly after the launches found that Fable’s defensive measures could be bypassed and that the system could surface vulnerabilities the restrictions should have blocked. Andy Jassy, Amazon’s CEO, informed Scott Bessent, the Treasury secretary, about the security concerns.
Independent experts later noted that Fable had not been fully compromised and therefore could not immediately have been turned into a tool for actual cyberattacks. Nonetheless, the U.S. government acted quickly: Sean Cairncross, the administration’s cybersecurity director, issued an ultimatum to Anthropic CEO Dario Amodei, demanding that the model be withdrawn within ninety minutes or that foreign access would be blocked.
President Trump assigned Howard Lutnick to handle the matter; Lutnick sent Anthropic an official letter notifying the company that restrictions on foreign users would take effect. The measures also apply to foreign nationals working in the United States, affecting Anthropic’s own international research staff.
Effects and implications
The incident deepened existing friction between the government and technology firms, which had clashed previously over topics such as military uses of AI, semiconductor exports to China, and Anthropic’s political ties. In early June, President Trump issued an executive order requiring developers to provide government access to new models 30 days before any public demonstrations.
Mythos has so far identified more than ten thousand bugs overall. Carlini highlighted the Ghost case to illustrate the risk: after a vulnerability was reported and patched, attackers analyzed the update and reversed the flaw, leading to the compromise of more than seven hundred websites by April.
Carlini warned it may be only a matter of months before other AI models reach Mythos’s capability level. In economies that rely on large amounts of relatively untested software infrastructure, models like Mythos can surface unprecedented numbers of security flaws, potentially shifting the long-standing balance between offensive and defensive sides of cybersecurity.
Current status
Anthropic has imposed a temporary shutdown of Mythos, and U.S. government restrictions bar foreign access to the model. Discussions between the company and officials are ongoing; government officials say Anthropic should acknowledge mistakes made when introducing Fable and improve communication with the White House.
(Additional developments and related reporting were cited in the source coverage.)
