Anthropic details Fable 5 cyber safeguards and a shared jailbreak severity scale

On July 2, Anthropic published a detailed look at the safeguards behind Claude Fable 5, the high-capability model it redeployed on July 1 after a brief US export-control suspension. The post matters to anyone building with AI coding tools, because it spells out exactly how a frontier model decides which security-related requests it will help with — and which it won't.

Four tiers instead of a blanket block

Rather than refusing anything that touches security, Anthropic says Fable 5 runs requests through classifiers that sort them into four categories:

• Prohibited use — activities with little-to-no defensive value, which are blocked.
• High-risk dual use — techniques widely abused by attackers that also have legitimate uses; these are blocked pending stronger authorization controls.
• Low-risk dual use — mostly defensive work, monitored and occasionally blocked.
• Benign use — everyday work that causes no harm, allowed with monitoring.

The practical upshot for developers: routine tasks like secure coding, patch management, log analysis, and incident response fall in the benign or low-risk buckets and are meant to work normally. Anthropic notes the classifiers err on the side of caution, so occasional false blocks on legitimate security work can still happen.

A severity scale for jailbreaks

The more novel piece is a proposed Cyber Jailbreak Severity (CJS) scale, developed with Glasswing partners including Amazon, Microsoft, and Google. A "jailbreak" is a prompt trick that gets a model to bypass its own safeguards; today the industry has no shared way to say how bad a given one is.

The CJS scale runs from CJS-0 (Informational) to CJS-4 (Critical) and rates each jailbreak on four axes:

• Capability gain — how far beyond existing tools it takes an attacker.
• Breadth of capability — how many distinct offensive tasks the same trick unlocks.
• Ease of weaponization — the effort needed to actually deploy it.
• Discoverability — how easily a threat actor could find the technique.

The idea mirrors how CVSS scores software vulnerabilities: a common vocabulary lets vendors and researchers triage consistently instead of arguing over whether a given exploit is "serious."

Inviting researchers in

Alongside the framework, Anthropic launched a HackerOne program where security researchers can submit cyber jailbreaks they discover in Fable 5 for review. The company says the classifier it deployed at redeployment blocks more than 99% of attempts to reproduce the specific jailbreak that prompted the earlier pause.

Why it matters for people learning to build

If you're learning to ship software with AI help, this is a useful window into how the tools reason about risk. Two takeaways: first, defensive and educational security work is explicitly supported, so you shouldn't expect a capable assistant to refuse ordinary tasks like writing input validation or reviewing your own code for flaws. Second, model providers are moving toward shared, measurable safety standards rather than one-off policies — a sign the ecosystem is maturing, and a framework other labs may adopt.

Anthropic is clear the CJS scale is an early draft meant to start an industry conversation, not a finished standard. But it's a concrete step toward treating AI safety incidents with the same rigor the security world already applies to software bugs.