Claude Fable 5 Lasted Three Days: A System Prompt Leak, Degraded Answers and a US Directive

Anthropic released Claude Fable 5 and the closed Mythos 5 on June 9, and on June 12 it pulled access to both. In the three days between release and withdrawal, more happened than some models see in a year: a system prompt leak, covert degradation of answers, a change to the data retention policy and a directive from the US government.

Launch and a sudden rollback

Fable 5 was positioned as a significant step forward for the Claude lineup: improved reasoning, more precise instruction following and deep contextual analysis. Mythos 5, a version with extended capabilities, was distributed only in a closed environment to select partners.

Launching two versions at once, one public and one closed, is unusual in itself: Anthropic normally releases a single model and expands access gradually.

Three days after release, the company suspended access to both models without explanation. The official statement only confirmed the suspension and did not give reasons. Even emergency rollbacks usually come with a short public explanation; this one did not. The lack of comment only fueled discussion and speculation.

The prompt leak and covert degradation

The researcher Pliny published a detailed post on X with Fable 5's alleged system prompt and put the full archive on GitHub. According to him, the model's safeguards were bypassed using multi-step prompt injection. Anthropic has not confirmed that the published material is authentic, but there was no official denial either, which many read as tacit agreement.

At the same time, a far more serious problem surfaced in the AI community: the company allegedly degraded the quality of answers for a number of AI researchers in secret. The practice reportedly meant that certain accounts, likely those actively probing the model's defenses, received noticeably worse answers than ordinary users. For the community that does independent evaluation of AI systems, this is a serious blow: if a model deliberately degrades for those who test it, no external benchmark can be considered reliable.

"If

Anthropic can deliberately degrade the model for certain users, every public test loses credibility," wrote one AI benchmarker.

The dispute over ways to bypass safeguards went beyond the academic: several independent teams published demonstrations of successful jailbreak scenarios against Fable 5, increasing pressure on the company to explain itself.

Data, tooling and regulation

Amid the scandals, Anthropic quietly updated its privacy policy: conversations with Claude are now stored for up to 30 days by default, instead of a shorter period. The change came without an official announcement. Outside observers noticed it, and it only added to the criticism of a company already accused of opacity.

On the developer side, Anthropic kept developing Claude Code during the same period. Key tooling updates include:

• Expanded support for agent chains
• New hook types (CwdChanged, FileChanged, TaskCreated)
• Improved handling of large monorepos
• Support for running subagents in parallel
• An updated permissions management interface

At the end of the period, a US government directive appeared that affects the use of Claude in government agencies. The details of the document have not yet been fully disclosed, but a regulatory act that names a specific commercial model is rare. It is a signal that oversight of powerful AI systems is moving from debate to practice.

What this means

In three days, Anthropic went from a loud release to a full withdrawal of the model, with stops at a leaked system prompt, a scandal over degraded answers and a quiet change to data retention terms. Each episode can be explained on its own, but together they form a pattern of opacity. As AI systems grow more capable, regulators and users will demand greater clarity, and the story of Fable 5 shows plainly what the lack of it costs.