Anthropic’s Claude Fable 5 returned on July 1 after US export controls were lifted, but the model had been paused because researchers found a way to bypass its safety guardrails. The fix is a new safety classifier that detects the specific bypass technique in over 99 % of cases. When the classifier flags a request, the system does not refuse outright; it routes the query to Claude Opus 4.8, allowing the user to continue working while the unsafe content is blocked. This fallback keeps workflows uninterrupted and gives developers a clear path to handle flagged prompts without redesigning their calls.
The incident also highlighted a missing industry standard for judging how dangerous a jailbreak is. Anthropic, together with partners, is proposing a four‑criteria framework: capability gain, breadth of gain, ease of weaponization, and discoverability. Each criterion is scored from low to high, and the combined score determines a response tier—from simple logging for minor issues to immediate mitigation and government notification for severe cases. Teams can adopt this framework to prioritize which vulnerabilities need urgent patches and which can be monitored over time.
During the pause, rivals such as GLM‑5.2 appeared, offering lower cost but fewer safeguards. For organizations that need strong cyber protections, Fable 5 remains the best choice because its classifier adds a safety margin that blocks harmful code generation while still letting legitimate debugging and coding pass through. Users should update their API calls to specify “claude-fable-5” and rely on the built‑in fallback to Opus 4.8 when needed. Monitoring the classifier’s alerts and applying the jailbreak severity score will help maintain both productivity and security.
#AI #Product #MachineLearning #Safety #LLM #EnterpriseAI