Anthropic has disputed allegations of a prompt-based jailbreak affecting its recently launched Claude Fable 5 AI model, underscoring the robustness of the advanced classifier system and extensive red-teaming efforts that underpinned the model’s deployment.
Claude Fable 5 became generally available on Tuesday, when Anthropic introduced it as a powerful Mythos-class AI model with safeguards that restrict its use in high-risk domains such as cybersecurity, where Mythos has proved particularly potent.
In sensitive areas such as cybersecurity, where it could be abused to develop exploits, and biology, where it could be leveraged to develop bioweapons and chemical weapons, the model automatically falls back to the less capable Claude Opus 4.8.
Anthropic said it conducted extensive internal and external red-teaming to ensure that Fable 5 cannot be easily jailbroken.
However, shortly after its release, an individual with the online moniker Pliny the Liberator, who is known for AI jailbreaks, claimed to have “liberated” Fable 5 by circumventing its restrictive safety layer.
The hacker said in a post on X that they used sophisticated multi-agent prompting methods, successfully eliciting useful information on sensitive topics, including cybersecurity, chemistry, psychological manipulation, and explosives.
Pliny the Liberator has published several screenshots to support the claims and released what is allegedly the Fable 5 internal system prompt, which contains instructions that define its personality, safety classifiers, fallback behaviors, tone guidelines, and refusal logic.
Contacted by SecurityWeek, an Anthropic spokesperson said the AI researcher’s post does not demonstrate a jailbreak of Fable 5’s safety systems.
The company explained that true jailbreaks would need to bypass its core safeguards and deliver meaningful assistance toward high-risk activities such as bioweapons development or sophisticated cyberattacks.
Instead, the demonstrated approach relies on coaxing the model to continue responding despite its conversational refusals, which is a well-known and longstanding limitation present in nearly all large language models.
Anthropic emphasizes that its strongest protections against the most dangerous risks are enforced by independent classifier systems that operate separately from the model itself, meaning that overcoming the model’s refusals does not disable these critical safeguards.
After examining the examples shared by the researcher, the company determined that some outputs were not produced by Fable 5 at all, while those that were contained only general information already available in public sources, offering no meaningful uplift for real-world harm.
A wider review of recent usage found no evidence of their safeguards being successfully circumvented to generate genuinely dangerous content, Anthropic said.
Related: After AI Reaches Production: 12 Ways Security Teams Can Take Control
Related: Claude Mythos Turns N-Days Into N-Hours With Rapid Exploit Creation
