Claude Fable 5 lasted three days: system prompt breach, degradation, and US directive

Anthropic released Claude Fable 5 and the closed Mythos 5 on June ninth — and by the twelfth had withdrawn access to both versions. Within three days between release and withdrawal, more happened than many a model accumulates in a year: system prompt compromise, hidden degradation of responses, change in data retention policy, and a directive from the U.S. government.

Launch and sudden rollback

Fable 5 was positioned as a significant step forward in the Claude lineup: improved reasoning, more precise instruction following, deep contextual analysis. Mythos 5 — a version with expanded capabilities — was distributed only in a closed circuit to select partners. The simultaneous launch of two versions — public and closed — was itself an atypical move: Anthropic usually releases one model with gradual access expansion. Three days after the release, the company suspended access to both models without explanation. The official statement merely confirmed the fact of suspension, without stating reasons. Even emergency rollbacks are typically accompanied by a brief public explanation — here there was none. The absence of comments only intensified waves of discussion and speculation.

Prompt compromise and hidden degradation

Researcher Pliny published a detailed post on X with the alleged system prompt of Fable 5 and placed the full archive on GitHub. According to him, the model's protective mechanisms were bypassed using multi-step prompt injection methods. Anthropic neither confirmed nor officially denied the authenticity of the published data, which many interpreted as tacit consent.

In parallel, the AI community uncovered a far more serious problem: the company allegedly secretly degraded response quality for a number of AI researchers. This refers to a practice in which certain accounts — likely those actively testing the model's protections — received noticeably worse responses than regular users. For a community engaged in independent evaluation of AI systems, this is a serious blow: if a model deliberately degrades for those who check it, no external benchmark can be considered reliable.

"If

Anthropic can deliberately degrade the model for specific users, any public test loses trust," wrote one AI benchmarker.

The dispute over protection bypass methods went beyond academic scope: several independent teams published demonstrations of successful jailbreak scenarios with Fable 5, which intensified pressure on the company for explanations.

Data, tools, 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 occurred without official announcement — independent observers noticed it, and this only added criticism against a company already accused of opacity. On the development front, Anthropic continued advancing Claude Code during the same period. Among key toolkit updates:

• Extended support for agent chains
• New hook types (CwdChanged, FileChanged, TaskCreated)
• Improved work with large monorepos
• Support for parallel subagent execution
• Updated permissions management interface

At the end of the period, a directive from the U.S. government emerged, affecting the use of Claude in government structures. Details of the document have not yet been fully disclosed, but the mere appearance of a regulatory act mentioning a specific commercial model is rare. This is a signal: oversight of powerful AI systems is shifting from discussion to practice.

What this means

In three days, Anthropic traveled from a major release to a complete model withdrawal — with stops at the compromised system prompt, the response degradation scandal, and the quiet change in data retention terms. Each episode individually is explicable, but together they form a pattern of opacity. As the capabilities of AI systems grow, regulators and users will demand greater clarity — the Fable 5 story starkly demonstrates what results from its absence.