More details on Fable 5’s cyber safeguards and our jailbreak framework
Key Points
- Fable 5 redeployed globally
- Four classifier categories; larger safety margin
- Draft jailbreak severity framework + HackerOne
Summary
Claude Fable 5 has been re-deployed globally. This note summarizes the model’s cybersecurity safeguards (the safety classifiers), how they categorize cybersecurity-related prompts, and an early draft AI jailbreak severity framework developed with partners. It explains what the classifiers are intended to block, allow, or monitor, and describes practical steps for security researchers and engineers interacting with Fable 5.
Key Points
- Classifier categories: prohibited use (blocked), high-risk dual use (blocked until stronger controls), low-risk dual use (monitored; sometimes blocked by a safety margin), and benign use (allowed, with monitoring).
- Safety margin: Fable 5’s safety margin was intentionally increased versus prior models, causing more conservative blocking (higher false positives) to reduce risk of harmful outputs.
- Prohibited examples: destructive malware, cyber-physical sabotage, defense evasion, command-and-control, exfiltration, malware development and delivery, internet backbone attacks — these are intended to be fully blocked.
- High-risk dual use examples: unauthorized access, exploitation, privilege escalation, lateral movement, exploit weaponization, ICS/OT/telecom/financial attack techniques — blocked pending stricter access controls.
- Low-risk dual use examples: OSINT, public scanning, vulnerability identification that many other tools already perform, and cryptographic protocol testing — generally allowed but some requests fall into the safety margin and may be blocked.
- Benign use: core defensive/IT activities are typically allowed; any blocks are likely false positives from the safety margin.
- Broader safeguards: classifiers operate alongside access controls, model safety training, and offline monitoring.
- Jailbreak framework: an early draft severity taxonomy to standardize how developers, governments, and researchers describe jailbreak risk; feedback is requested to refine it.
- How to engage: submit potential cyber jailbreaks via the new HackerOne program and send feedback to cyber-safeguards@anthropic.com.
Practical guidance for engineers
- Expect stricter blocking for high-risk and prohibited prompts; design tests and automation to handle classifier rejections (clear error handling and human review paths).
- When researching model behavior or potential jailbreaks, use the HackerOne program rather than posting exploits publicly; include context and reproduction steps.
- If you rely on vulnerability-finding workflows, validate whether Fable 5’s classifier behaviour meets your needs (some high-uplift vulnerability finding is intentionally blocked).
What’s next
- Anthropic may adjust classifier boundaries and the safety margin over time based on real-world feedback and security research. The jailbreak severity framework is a work in progress and they seek community input.