Anthropic details Claude Fable 5 cyber safeguards

Anthropic has released new details about the cybersecurity safeguards protecting Claude Fable 5 and proposed a draft framework for assessing the severity of AI cyber jailbreaks.

The company said Fable 5 has now been redeployed globally and is available to all users. The update focuses on two areas: safety classifiers designed to detect and block dangerous cybersecurity activity, and a proposed framework for grading the severity of cyber-related AI jailbreaks.

Anthropic said Fable 5 uses safety classifiers to distinguish between four categories of cybersecurity activity: prohibited use, high-risk dual use, low-risk dual use and benign use. Prohibited activities include destructive cyberattacks, cyber-physical sabotage, defence evasion, command-and-control infrastructure, data exfiltration, malware development and attacks targeting internet backbone systems.

High-risk dual-use activities include penetration testing, red teaming, privilege escalation, exploit development, virtual machine escapes and advanced vulnerability discovery. Anthropic said these activities will remain restricted for Fable 5 until stronger access controls for trusted users are in place.

The proposed Cyber Jailbreak Severity (CJS) framework introduces a scale ranging from CJS-0 to CJS-4, assessing jailbreaks according to capability gain, breadth of capability gain, ease of weaponisation and discoverability.

Anthropic said the framework is intended to give AI developers, governments and security researchers a common vocabulary for describing cyber jailbreak risks. The company is seeking public feedback on the draft and has launched a HackerOne programme inviting researchers to submit potential cyber jailbreaks affecting Fable 5.

Why does it matter?

As AI models become more capable in cybersecurity, assessing jailbreaks consistently is becoming increasingly important. A common severity framework could help AI developers, researchers, and governments compare vulnerabilities, prioritise responses and communicate risks using shared criteria rather than ad hoc descriptions.

The proposal also reflects a broader shift toward treating frontier AI safety as a collaborative effort. By publishing its methodology and inviting external researchers to test Fable 5 through a coordinated disclosure programme, Anthropic is encouraging greater transparency and independent scrutiny of AI security safeguards.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!