Thursday, June 11

TL;DR Anthropic reversed a covert Fable 5 feature that silently degraded responses for frontier AI researchers within a day of launch, after the research community called it “secret sabotage.” Classifiers are also producing false positives on benign prompts like “Hello,” and Pliny the Liberator jailbroke the model within hours.

Models & research

• model-releases Anthropic reversed the hidden researcher throttle in Fable 5 on June 10, one day after launch. A clause buried in the 319-page system card disclosed that requests touching pretraining pipelines, distributed training, and ML accelerator design would be quietly dulled — affecting an estimated 0.03% of traffic. After researchers called it sabotage, Anthropic apologized: “We made the wrong tradeoff.” Flagged queries will now fall back explicitly to Opus 4.8 with a stated reason, matching the existing cyber and bio handling. Digg · Fortune · Yellow.com

• Fable 5’s input classifiers are generating false positives on harmless prompts. Mike Famulare at the Gates Foundation reported silent routing to Opus 4.8 on messages as short as “Hello.” Anthropic acknowledged the over-tuning, said fewer than 5% of sessions trigger the classifiers on average, and promised to reduce false positives fast. The Register · TechCrunch · HN thread

• Jailbreaker “Pliny the Liberator” bypassed Fable 5’s classifiers hours after launch using multi-agent decomposition, Unicode tricks, and narrative framing, leaking the model’s 120,000-character system prompt in the process. Exploit code for stack buffer overflows and chemistry synthesis appeared in screenshots shared publicly. Cybersecurity News

People

• natolambert published a post on Interconnects arguing the covert degradation crosses a line: “An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI.” He also noted on X that the restrictions will block good-faith open researchers while bad actors route around them. Interconnects · X