Mythos Returns: Anthropic’s Tests Show Fable 5 Wasn’t Uniquely Risky

Anthropic says internal testing found Claude Fable 5 posed no unique cybersecurity danger, as Claude Mythos 5 returns globally on July 2.

The admission accompanies Fable 5’s global relaunch, capping an 18-day suspension triggered by US export controls on June 12. Anthropic tested rival models to gauge the real threat behind the restriction.

Why Anthropic Suspended Fable 5

Fable 5 and Mythos 5 launched June 9, sharing the same core model with the former open to the public. Mythos 5 stayed limited to a small number of trusted Project Glasswing partners for defensive cybersecurity work.

The export controls arrived after Amazon researchers found a way to bypass Fable 5’s safeguards. The technique prompted the model to identify software vulnerabilities and, in one case, demonstrate an exploit.

Claude Fable 5 will be available again globally tomorrow.After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding…

— Anthropic (@AnthropicAI) July 1, 2026

Anthropic’s tests found that Claude Opus 4.8, GPT-5.5, and Kimi K2.7 could identify the same vulnerabilities Fable 5 flagged in the Amazon report. Every model tested could reproduce the single exploit demonstration too.

The finding suggests the directive targeted a gap shared across the industry, not a Fable-specific threat. Anthropic still built a stronger classifier to block the technique, which now also flags more routine coding and debugging requests.

How The Guardrails Actually Work

Fable 5 launched with the strongest safety margin Anthropic has built into any model. Its classifiers block requests that look even slightly risky, not just the clearly harmful ones. The new classifier trained after the Amazon report blocks the reported bypass in over 99% of cases, according to Anthropic. Blocked requests now reroute automatically to Opus 4.8.

That safety margin comes at a cost. Anthropic acknowledges the classifier flags more benign coding and debugging requests, and says it will keep tuning it to cut false positives. Mythos 5, which carries fewer of these guardrails, returned only for Mythos 5 institutions cleared by the government on June 26.

Anthropic’s own data raises a harder question. If weaker models can already do what Fable 5 was banned for, what standard will regulators apply the next time a frontier model launches?