Anthropic is disputing reports that its newly released AI model, Claude Fable 5, was successfully jailbroken within 48 hours of its June 9 debut. The company asserts that over 1,000 hours of external bug bounty testing failed to produce a universal bypass of its safety protocols.
Allegations surfaced claiming researchers used novel multi-agent techniques to circumvent guardrails and leak a 120,000-character system prompt. Anthropic denies these claims but has not yet issued a detailed technical rebuttal addressing the specific attack vectors.
The model employs a layered safety architecture that routes high-risk queries to a more restricted system. However, this design doubles token consumption compared to previous versions and enforces a mandatory 30-day data retention policy, potentially complicating adoption for regulated industries.
If verified, the alleged prompt leak could expose Anthropic’s proprietary alignment logic to competitors. Furthermore, persistent security vulnerabilities in frontier models may accelerate regulatory scrutiny and increase compliance costs across the artificial intelligence sector.