Anthropic Disputes Claim That Claude Fable 5 Was Jailbroken

Summary

A researcher claimed to jailbreak Claude Fable 5 using multi-agent prompts. Anthropic disputed that the examples bypassed core safeguards, saying some outputs were not from the model and others offered no meaningful harmful capability.

Key points

Jailbreak researcher Pliny the Liberator claimed multi-agent prompts elicited responses on sensitive topics, including cybersecurity and explosives.
Anthropic said the examples did not demonstrate bypassing Fable 5’s core safeguards; some outputs were not produced by the model, and others contained only public information.
Anthropic distinguishes coaxing the model to continue after a conversational refusal from defeating separate classifier systems that enforce safeguards.
The company said a wider review found no evidence that safeguards had been bypassed to generate genuinely dangerous content.
Anthropic said Fable 5 falls back to Claude Opus 4.8 for sensitive areas such as cybersecurity and biology.

Article Details

Event Type: Dispute over an alleged prompt-based jailbreak of Anthropic’s Claude Fable 5 AI model.
Impact: Pliny the Liberator claimed to have bypassed the model’s safety layer and shared examples and an allegedly leaked system prompt. Anthropic disputed that the examples demonstrated a jailbreak of Fable 5’s core safeguards, stating that some outputs were not produced by the model and the others provided only general information already publicly available. Anthropic said its separate classifier safeguards remained in place and that a review found no evidence of those safeguards being successfully circumvented to generate genuinely dangerous content.

People

Pliny the LiberatorThe individual who claimed to have circumvented Fable 5’s restrictive safety layer, and published screenshots and an allegedly leaked internal system prompt. Anthropic disputed that the examples demonstrated a jailbreak of the model’s core safeguards.

Vendors

AnthropicAnthropic 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

Products

Claude Fable 5Anthropic 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 Claude Opus 4.8to develop bioweapons and chemical weapons, the model automatically falls back to the less capable Claude Opus 4.8.MythosClaude 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