Anthropic allows Fable 5.1 for vuln scanning, but prompt triggers downgrade
AccBalanced · x · 2026-09-02
Anthropic's blog states that Fable 5.1 is now permitted for identifying software vulnerabilities. However, testing shows that when prompted to "Review this code for vulnerabilities," the model triggers a safety refusal and downgrades to Opus 4.8, creating a contradiction between the announced policy and actual behavior.
More from Fun
- Five-year-old AI-generated sci-fi short 'Blightsight' resurfaces — generativist · 2026-09-02
- Midjourney founder David Holz: rent 8x GB300 racks for a weekend to attack 3000 open math conjectures — ethanniser · 2026-09-02
- Fable 5.1 Helps Crack 373-Year-Old Cipher; Solution Deemed Simple — ethanniser · 2026-09-02
- Google's "best" image model generates a student as a mascot — burkov · 2026-09-02
- "Time Freeze" visual effect powered by Atlas — gowthami_s · 2026-09-02
- Claude vs. ChatGPT comparison using leaked Fable 5.1 system prompt — seanmcdonaldxyz · 2026-09-02