Fable 5's Cyber Safeguards: 4 Categories, 1 Jailbreak Scale
Fable 5 re-deployed globally on July 2, 2026 with published rules for its cybersecurity classifiers and a 5-level jailbreak severity scale. This lesson shows where the lines are drawn and how to frame security work so Fable 5 helps instead of blocks.
Fable 5 re-deployed globally on July 2, 2026 — this time with a published rulebook for its cybersecurity classifiers.
The key fact: Fable 5 does not block all security content. It sorts every request into one of four categories. Your category determines whether Fable 5 helps freely, monitors and sometimes blocks, or refuses entirely.
Knowing the four categories tells you exactly what to ask, how to ask it, and what to do when a request gets blocked.
Prohibited use — Always blocked. Ransomware, malware development, defense evasion (AV/EDR bypass), command-and-control infrastructure, data exfiltration from victim devices, and backbone attacks like BGP hijacking. The asymmetry is the rule: if attackers gain far more than defenders, it's prohibited.
High-risk dual use — Blocked for now. Penetration testing, exploit development, privilege escalation, lateral movement, and zero-click vulnerability research. All legitimate security work — but Fable 5 blocks this category until better controls exist to verify who's asking and under what authorization.
Low-risk dual use — Mostly allowed, sometimes blocked. OSINT on public systems, vulnerability identification that other tools can already do, SSL/TLS testing. Fable 5 applies a "safety margin" here: it blocks a wider band of ambiguous requests than previous models, accepting more false positives in exchange for greater confidence that nothing harmful slips through.
Benign use — Always allowed. Secure coding, log analysis, incident response, malware reverse engineering, patch management, security certifications, and disaster planning. If Fable 5 blocks a clearly legitimate security question, it has likely landed in the safety margin — not a hard block, not a jailbreak opportunity.
The Cyber Jailbreak Severity (CJS) scale is a new framework Anthropic developed with its Glasswing partners to grade how dangerous a given jailbreak actually is. Jailbreaks are prompt techniques that bypass a model's safeguards. The CJS scale runs from CJS-0 (None/Informational) to CJS-4 (Critical), with bands that are exponential — each step is several times more serious than the last.
Four axes drive the score: capability gain (how far beyond existing tools does the jailbreak take an attacker?), breadth (does it work on one target or many?), ease of weaponization (how much effort to run an attack?), and discoverability (how easily can a threat actor find it?). A jailbreak scoring 0 on capability gain is automatically CJS-0 — if existing public tools can do the same thing, there is no new risk. Anthropic launched a HackerOne program for researchers to submit Fable 5 jailbreaks, and is collecting feedback at cyber-safeguards@anthropic.com.
What you can do now
Benign security work — log analysis, incident response, reverse engineering, secure coding, patch management — use Fable 5 directly.
High-risk dual use work — pentest steps, exploit dev — Fable 5 blocks this category for now. Use specialized tools for those steps; use Fable 5 for documentation and remediation guidance.
Safety margin blocks — if a clearly legitimate question gets refused, reframe it toward the defensive outcome. The classifier responds to context signals.
CJS-0 through CJS-4 gives you and your security team a shared vocabulary for discussing jailbreak risk with management and government stakeholders.