Anthropic Details Claude Fable 5 Cyber Safeguards and AI Jailbreak Severity Framework

Anthropic has published new technical details on the cybersecurity safeguards protecting Claude Fable 5, which is now globally available following its redeployment. The disclosure covers two major initiatives: a breakdown of the model’s safety classifiers, and an early-draft framework for scoring the severity of AI jailbreaks developed in partnership with Anthropic’s Glasswing collaborators. Claude Fable […] The post Anthropic Details Claude Fable 5 Cyber Safeguards and AI Jailbreak Severity Framework appeared first on Cyber Security News.

Anthropic has published new technical details on the cybersecurity safeguards protecting Claude Fable 5, which is now globally available following its redeployment. The disclosure covers two major initiatives: a breakdown of the model’s safety classifiers, and an early-draft framework for scoring the severity of AI jailbreaks developed in partnership with Anthropic’s Glasswing collaborators. Claude Fable 5 Cyber Safeguards and AI Jailbreak Severity Framework Because cybersecurity capabilities are inherently dual-use, Anthropic designed Fable 5’s classifiers to sort requests into four tiers rather than issuing blanket blocks: Prohibited use: Ransomware, wipers, cyber-physical sabotage (ICS/SCADA targeting), defense evasion, malware development/delivery, C2 infrastructure, and internet backbone attacks (BGP hijacking, DNS/CA compromise). Always blocked due to high harm-to-benefit asymmetry. High-risk dual use: Penetration testing, exploitation, privilege escalation, exploit development, and VM/container escapes. Blocked by default pending better authorization controls, alongside efforts to prevent “high-uplift” vulnerability discovery flaws that only top-tier models or experts could otherwise find. Low-risk dual use: OSINT, public vulnerability scanning, and cryptographic protocol testing. Mostly allowed, with selective blocking as part of a deliberate “safety margin.” Benign use: Secure coding, debugging, patch management, incident response, and security education. Allowed by default; any blocks are treated as false positives. Anthropic notes that Fable 5’s safety margin was set wider than in prior models, trading a higher false-positive rate for stronger assurance against jailbreaks slipping through. Classifiers work alongside access controls, model training, and offline monitoring as layered defenses. The second component of the disclosure is a proposed five-band severity scale, ranging from CJS-0 (Informational) to CJS-4 (Critical), designed to give AI developers and governments shared language for discussing jailbreak risk. Scores derive from four axes, each rated numerically and summed. Capability gain measures how far a jailbreak advances an attacker beyond existing tools, while breadth captures how many distinct attack types or vulnerability classes the technique applies to. Ease of weaponization reflects how much effort is needed to convert the technique into a working exploit, and discoverability measures how readily a threat actor could find the technique independently, whether through public disclosure or reasonable red-team effort. The combined score sets a severity “floor,” which Anthropic can raise, but never lower, based on discretionary factors such as unpatchable root causes or compounding risk when a jailbreak links with other open findings. Anthropic is soliciting feedback at cyber-safeguards@anthropic.com and has launched a dedicated HackerOne bug bounty program for researchers to report cyber jailbreaks found in Fable 5. The company frames this as a starting point for cross-industry standardization, explicitly excluding non-cybersecurity jailbreaks, such as system-prompt extraction, from the framework’s scope since Anthropic already publishes those prompts itself. The move reflects growing pressure on frontier AI labs to formalize dual-use risk management as models grow more capable in offensive security domains, an area regulators and CISOs alike are watching closely heading into the second half of 2026. Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google. The post Anthropic Details Claude Fable 5 Cyber Safeguards and AI Jailbreak Severity Framework appeared first on Cyber Security News.