Anthropic releases Claude Fable 5 with security guardrails amid jailbreak risks
Anthropic released Claude Fable 5, a publicly available version of its powerful Mythos-class AI model, with built-in security guardrails to prevent misuse in cybersecurity, biology, and chemistry. Security researchers reportedly made progress toward jailbreaking the guardrails, and Microsoft restricted the model's use due to data retention concerns.
What happened
Anthropic publicly released Claude Fable 5, a model of the Mythos class, which earlier drew controversy because it could rapidly identify vulnerabilities in software security. The model was made publicly available after Anthropic had previously restricted access to a preview version to a limited number of organizations.
Anthropic implemented security guardrails in Fable 5 that fall back to Claude Opus 4.8 responses for requests related to cybersecurity, biology, chemistry, and model copying. External security firms found no complete bypasses of these guardrails, though Anthropic acknowledged such bypasses might exist. The UK AI Security Institute made progress toward a jailbreak during testing, but Anthropic stated such attacks would require significant effort and cost to execute at scale.
Reported impact
Anthropic stated that without security measures, Fable 5's capabilities in cybersecurity could be misused to cause serious harm. Microsoft restricted use of Claude Fable 5 due to data retention concerns.
Index analysis
The release of a model with powerful capabilities for identifying software vulnerabilities presents risks that jailbreak techniques could circumvent the implemented guardrails, potentially exposing its capabilities for malicious use in cybersecurity attacks.
Governance controls
Anthropic subjected the security guardrails to extensive testing and had external security firms audit them. The guardrails route sensitive requests through a less capable model to limit capability exposure.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.