The Fable 5 'Jailbreak' Behind the Shutdown — and Anthropic's Pushback

SecurityWeek·high signal

The bypass that triggered the federal order was claimed by 'Pliny the Liberator,' using multi-agent prompting plus Unicode/homoglyph and Cyrillic substitutions, long-context reference tracking, fiction/academic framing, and decomposed requests to coax responses on cyber, chemistry, and explosives topics. Anthropic disputes it is a true jailbreak, saying its review found only minor, previously known vulnerabilities already reachable with other public models, and that the technique merely coaxes continued responses despite refusals — a well-known limitation of nearly all LLMs. For builders, it's a live case study in how guardrail-bypass disclosures can carry regulatory, not just security, consequences.