Amazon CEO's Security Warnings Preceded Anthropic Model Export Ban
Andy Jassy flagged Claude Fable 5 vulnerabilities to U.S. officials before government restricted the model's distribution.
Last verified:
BLUF
Amazon CEO Andy Jassy conveyed security vulnerabilities in Anthropic’s Claude Fable 5 to U.S. officials, including Treasury Secretary Scott Bessent, before the government enacted an export control ban on both Fable 5 and Mythos 5 on Friday, June 12. The Wall Street Journal and Reuters both reported that Amazon researchers had identified methods to exploit the model for cyberattack planning, triggering the intervention.
Amazon’s Pre-Ban Warning Channel
According to The Wall Street Journal, Jassy informed Treasury officials and other government personnel that Amazon researchers had found ways to extract potentially harmful intelligence from Claude Fable 5—information with applicability to cyber operations. The Wall Street Journal and Reuters confirmed Amazon’s escalation of these findings through official channels, independent of any formal regulatory inquiry.
An Amazon spokesperson acknowledged the company’s engagement with government on security matters but declined to elaborate, stating it is “standard practice” to discuss potential risks without disclosing specifics. AWS customers subsequently lost access to both restricted models following the government action.
Anthropic’s Stated Position and Official Reaction
David Sacks, co-chair of the President’s Council of Advisors on Science and Technology and former AI czar, provided his own account of events. According to Sacks, an unnamed “highly credible trusted partner” (widely understood to reference Amazon) disclosed a jailbreak technique to federal officials. Sacks reported that the administration then asked Anthropic CEO Dario Amodei to patch the vulnerability or restrict deployment, a request Amodei allegedly declined.
This characterization differs markedly from Anthropic’s operational response—the company did restrict access worldwide rather than defend the models’ robustness. The gap between Sacks’s narrative and Anthropic’s action suggests either disagreement over remediation timelines or a shift in corporate calculus once the government imposed binding restrictions.
Why This Matters
For enterprise teams relying on Anthropic models in regulated sectors, this episode establishes that export bans can activate without advance notice or remediation windows. Compliance officers integrating Claude Fable or other frontier models must now assume that unpatched vulnerabilities disclosed through backchannels can trigger instantaneous access revocation, rather than orderly deprecation schedules. The precedent suggests that investor-government coordination on AI safety may bypass formal disclosure processes in future incidents.
Frequently Asked Questions
Why would Amazon flag security concerns about Anthropic's models if Amazon is a major investor?
Amazon holds both strategic investment and operational stakes—AWS customers rely on Anthropic model access. Reporting vulnerabilities protects both entities from downstream liability and regulatory exposure.
What does 'jailbreak' mean in David Sacks's statement?
A jailbreak is a technique that bypasses a model's safety guardrails to generate harmful outputs. Sacks's claim that Anthropic refused to address the vulnerability before deployment suggests the company chose speed over remediation.
Who is David Sacks and why does his account matter?
Sacks serves as co-chair of the President's Council of Advisors on Science and Technology (PCAS) and previously held the AI czar role. His insider position gives his narrative credibility in policy circles, though it remains one interpretation of events.