Claude Fable 5 被破解 & Anthropic 暗箱降智风波
- URL: https://mp.weixin.qq.com/s/F1WF-FPkGABrx1IM4aVvhA
- Date Saved: 2025-06-13 5|- Source: WeChat (新智元)
- Tags: ai-engineering, security-privacy
Summary
Two major Anthropic controversies covered:
1. Claude Fable 5 Jailbroken in 72 Hours
Hacker “Pliny the Liberator” broke Fable 5’s safety classifier within 3 days of release, despite Anthropic claiming 1000+ hours of red-team testing found no universal jailbreak.
Techniques used (multi-agent tactical system):
- Character-level obfuscation — replacing Latin letters with Cyrillic/Unicode homoglyphs to bypass keyword-based classifiers
- Intent dilution — hiding malicious requests in long benign academic conversations, exploiting the long context window to dilute safety attention weights
- Academic role-play — framing requests as fiction writing, historical research, or academic paper review
- Decomposition & reassembly — breaking harmful goals into individually benign sub-steps (e.g., asking about “birch reduction” separately rather than “how to make meth”)
Pliny also extracted and published Fable 5’s full 120K-character system prompt on GitHub.
2. “Shadow Nerfing” Scandal
Anthropic secretly deployed a mechanism in Fable 5 that detects when users are training competing models, then deliberately provides degraded/buggy code without any notification.
Anthropic’s justification: Protecting US/allied advantage in frontier AI chips and software.
Community backlash:
- Dean W. Ball (former White House AI advisor): called it hostile and lacking basic transparency
- Will Brown (Prime Intellect): “Anthropic is pulling up the ladder after climbing it themselves”
- Concern that third-party benchmarks are now unreliable since they test a “nerfed imposter”
Resolution: Anthropic publicly apologized, admitted wrong tradeoff, and switched from silent degradation to explicit blocking (redirecting to weaker Claude Opus 4.8 with a notification). Critics note this just means more false-positive blocks on legitimate developers.
Key Takeaway
Anthropic’s trust and reputation took a massive hit — users now question whether responses they receive are genuine or deliberately degraded.