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

Summary

Anthropic has released additional details on the safety classifiers deployed with Claude Fable 5, now available globally. These classifiers distinguish among four tiers of cybersecurity-related requests—prohibited, high-risk dual-use, low-risk dual-use, and benign—rather than attempting to block all security-related activity. The distinction reflects the dual-use nature of many capabilities, such as vulnerability discovery, which defenders rely on for legitimate assessments yet attackers can also exploit.

The prohibited category covers actions with limited defensive value and high potential for harm, including defense evasion and data exfiltration; these are blocked outright. High-risk dual-use activities, such as privilege escalation, lateral movement, and exploit development, are also blocked by default because they closely mirror offensive operations, even though they occur in authorized testing. A safety margin is applied around these boundaries so that prompts must appear unambiguously benign to pass, reducing the chance that harmful requests slip through. Low-risk dual-use prompts lean toward defensive tasks and are more likely to be permitted, while clearly benign activities such as routine code scanning or patch management are intended to remain available.

Alongside the classifier design, Anthropic has shared an early draft of a Cyber Jailbreak Severity (CJS) framework developed with Glasswing partners. The framework rates jailbreaks on a five-band scale from CJS-0 (informational) to CJS-4 (critical) according to four axes: the degree of capability gain beyond existing public tools, the breadth of targets or tasks affected, the ease of reproducing the jailbreak, and the level of expertise required to weaponize it. The bands are intended to be exponential, so that higher scores represent substantially greater real-world risk. The company invites feedback on both the classifier categories and the severity framework to support consistent discussion among developers, researchers, and policymakers.

Why it matters

Provides actionable, specific guidance on model-level cyber safeguards and a structured jailbreak evaluation rubric directly usable by product teams building or auditing AI systems, with clear discussion of dual-use risks and deployment trade-offs.