Talks between AI firm Anthropic and the White House have concluded without lifting export controls on the company's advanced AI models, despite Anthropic's insistence that concerns are overblown.
The administration is holding firm on its position, believing that Anthropic's Claude Fable 5 model can be manipulated to bypass its safeguards, effectively unlocking the powerful cybersecurity features of its underlying Mythos model. These concerns were raised last week, reportedly after Amazon CEO Andy Jassy contacted Treasury Secretary Scott Bessent, prompting an NSA review that supported the alleged vulnerabilities.
Working group meetings were held at the Commerce Department, involving representatives from Anthropic, including co-founder Tom Brown and head of external affairs Sarah Heck, alongside government researchers. Commerce Secretary Howard Lutnick participated via conference call.
Anthropic has pushed back, stating that the administration's concerns are exaggerated. The company previously released a blog post suggesting the risks were not as severe as portrayed. Some cybersecurity researchers echoed this sentiment in an open letter, arguing that the export controls hinder defenders and create market uncertainty without a justifiable threat.
The core of the disagreement lies in the interpretation of "jailbreaking" – the act of bypassing an AI's safety features. While Anthropic believes its Fable 5 model's guardrails are robust enough for public release, the government views them as potentially circumventable, especially given Mythos's advanced capabilities.
This situation has implications for the broader AI industry, with experts suggesting that labs may need to provide early access to advanced models and maintain proactive communication with the government regarding new releases. Anthropic has stated that both parties are working diligently to resolve the issue, though the Commerce Department indicated that a resolution would likely hinge on Anthropic fully addressing the jailbreak concerns.
