Feds freaked over Fable 5 after simple 'fix this code' prompt, not jailbreak

Article excerpt

Federal agencies grew alarmed over Fable 5, a large language model, after researchers demonstrated it could be manipulated with straightforward requests to fix buggy code, not through sophisticated jailbreak techniques. The vulnerability reportedly allowed the model to bypass safety constraints using innocent-seeming prompts rather than adversarial attacks. A researcher clarified that the method was simpler than many assumed, raising questions about how easily AI systems can be subverted through ordinary user interactions. The incident highlights the gap between theoretical security measures and real-world robustness.