Claude Fable 5 is set to make a global return on July 1, becoming accessible across Anthropic's platforms. Meanwhile, Claude Mythos 5, a version of the same core model with fewer safety restrictions, is being reintroduced exclusively to pre-approved US organisations, a decision made after thorough government review.
The initial suspension, which began on June 12, was a direct consequence of a significant cybersecurity incident. Researchers at Amazon identified a 'jailbreak' technique capable of bypassing Fable 5's embedded safeguards. This vulnerability allowed the model to identify software weaknesses and, critically, generate exploit code. The immediate imposition of export controls, which regulate access to technology by foreign nationals, prompted Anthropic to suspend access for all users. This preemptive measure was taken to avoid potential breaches, as real-time user nationality verification was not feasible.
The temporary outage had notable repercussions in the financial markets. Although Anthropic is not publicly traded, its valuation is tracked through a pre-IPO perpetual futures contract on the onchain exchange Hyperliquid. CoinDesk reported a 3.7% dip in this contract's value following the suspension in June, raising questions among investors about the potential timeline for a public listing.
In response to this episode, Anthropic is significantly enhancing its collaboration with the US government. This intensified partnership includes providing designated agencies with early access to upcoming frontier AI models and their associated safeguards before their public release. Furthermore, Anthropic will be sharing valuable threat intelligence and actively working towards establishing a unified security standard for AI developers across the industry.
Further demonstrating its commitment to safety, Anthropic is in the process of drafting a new framework alongside tech giants Amazon, Microsoft, and Google. The objective of this framework is to create a standardised system for rating the inherent danger of various jailbreak techniques. These proactive measures align with a broader executive order on AI security issued on June 2, signalling a tighter and more integrated relationship between AI developers and government oversight in the evolving landscape of artificial intelligence security.
