More details on Fable 5’s cyber safeguards and our jailbreak framework

Claude Fable 5 is back online everywhere, and Anthropic used the moment to pull back the curtain on something it usually keeps vague: the actual rules its safety classifiers follow when someone tries to use the model for hacking-adjacent tasks. Rather than a blanket ban on anything cyber-related, the company sorts requests into four buckets — prohibited, high-risk dual use, low-risk dual use, and benign — and treats each one differently depending on how lopsided the harm-to-benefit ratio looks.

The prohibited bucket is the clean block list: ransomware, wipers, malware development, command-and-control infrastructure, attacks on power or water systems, BGP hijacking, that sort of thing. Anthropic argues these offer attackers a lot and defenders comparatively little, even though some, like defense evasion, do show up in legitimate security work. High-risk dual use is messier — penetration testing, privilege escalation, exploit development, container escapes — activities that are completely normal for a red teamer with authorization but indistinguishable from an attack when the context is missing. For now, Fable 5 blocks these outright until Anthropic has better ways to verify who's asking.

Vulnerability finding gets its own carve-out worth noting. Anthropic isn't trying to stop Claude from finding bugs other tools can already find — that's core defensive work. What it wants to block is high-uplift discovery, meaning vulnerabilities that no other widely available model can locate, plus automatic exploit generation. The logic: if a jailbreak lets Fable uniquely surface bugs nothing else can find, that capability is too valuable to attackers to risk leaking out.

Low-risk dual use and benign categories get a lighter touch — things like OSINT gathering, TLS testing, patch management, log analysis, and secure coding are mostly allowed, though Anthropic admits it still blocks a chunk of low-risk requests as a deliberate safety margin, accepting some false positives to keep dangerous prompts from slipping through. For Fable 5 specifically, that margin was set wider than on earlier models.

Alongside the classifier breakdown, Anthropic published an early draft of a jailbreak severity framework developed with Glasswing, essentially a shared vocabulary for how bad a given jailbreak actually is, based on how far it pushes an attacker beyond existing tools and how broad the newly unblocked capabilities are. The company is inviting outside feedback via email and has opened a HackerOne program for researchers to report jailbreaks they find in Fable 5, framing the whole exercise as a starting point for industry and government to eventually agree on common terms rather than each lab defining risk on its own.