Anthropic Publishes Fable 5 Cyber Safeguard Details — New Jailbreak Framework
Anthropic published detailed technical documentation of Fable 5's cybersecurity architecture and jailbreak assessment framework after US AISI approval.
Introduction: Fable 5 Returns with Enhanced Security
On July 2, 2026, Anthropic officially published detailed technical documentation of Claude Fable 5's cybersecurity safeguard mechanisms. This came after the model received approval from the US AI Security Institute (US AISI) for redeployment. Fable 5 — Anthropic's newest and most powerful AI model — became globally available on July 1, and with it, the company introduced two important innovations: a detailed description of its cybersecurity classifiers and an entirely new framework for assessing the severity of AI jailbreaks.
This move is particularly significant given the ongoing global debate about AI safety. The release of transparent technical documentation represents a new standard for the industry, where AI companies are increasingly expected to not only build safe models but also explain their safety mechanisms in detail. For the Georgian tech community, this is an important signal: safety and transparency are becoming competitive advantages in the AI industry, not just regulatory requirements.
Fable 5's Cybersecurity Classifiers: Four-Tier Architecture
One of Anthropic's main challenges is that cybersecurity tools are often dual-use — the same capability can be used for both defensive and offensive purposes. For example, scanning code for vulnerabilities helps cybersecurity professionals, but the same ability could be used to prepare a cyberattack. This is the fundamental tension that Fable 5's safety architecture must navigate: enabling legitimate security research while preventing malicious applications.
This is why Fable 5's safety classifiers do not block all cybersecurity-related actions. Instead, they divide them into four categories:
- Prohibited use — actions capable of causing significant harm without defensive utility. This includes ransomware, wipers, DoS attacks, malware creation, C2 infrastructure, critical infrastructure sabotage, and others.
- High-risk dual use — actions widely used by cybercriminals but also having legitimate defensive applications. Examples: pentesting, exploit development, privilege escalation.
- Low-risk dual use — actions predominantly used for defensive purposes. Examples: OSINT, cryptographic protocol testing.
- Benign use — actions that pose minimal security risk. Examples: general coding assistance, documentation, education about cybersecurity concepts without providing active exploit code.
Cyber Jailbreak Severity (CJS) Framework — Industry's First Standard
Beyond classifiers, Anthropic introduced the Cyber Jailbreak Severity (CJS) framework — the industry's first standardized system for assessing the severity of AI jailbreak attempts in the cybersecurity domain. This addresses a critical problem: without a common language to describe how dangerous a jailbreak is, researchers, companies, and regulators cannot effectively communicate about threats.
The CJS framework rates jailbreak severity on a scale from CJS-0 (minimal risk) to CJS-4 (critical risk), similar to how the Common Vulnerability Scoring System (CVSS) works for software vulnerabilities. This parallel is intentional — Anthropic wants the AI safety industry to have the same kind of standardized severity assessment that the cybersecurity industry has used for decades.
The Four Axes of Jailbreak Severity Assessment
Each CJS rating is determined by assessing the jailbreak across four axes: Capability Gain (how much additional capability does the jailbreak unlock?), Breadth (how many users or systems are affected?), Weaponization Ease (how easy is it to turn the jailbreak into a real attack?), and Discoverability (how likely is it that others will find the same vulnerability independently?).
For Georgian security professionals and AI developers, this framework provides a practical tool: if you're evaluating an AI model for use in your business, you can now assess not just whether the model can be jailbroken, but how severe that jailbreak would be. This distinction is crucial for risk management in AI adoption.
The Glasswing Partnership and HackerOne Program
Anthropic partnered with Amazon (AWS), Microsoft (Azure), and Google (GCP) through what it calls Project Glasswing — a collaboration to ensure that Fable 5's cybersecurity capabilities don't create systemic risks across major cloud platforms. They also launched a dedicated HackerOne bug bounty program focused specifically on Fable 5 jailbreaks, with rewards of up to $50,000 for critical findings.