Claude Fable 5's cyber safety measures, revealed! Plus a new framework for measuring how dangerous a jailbreak is

Hello, it's Shii! Today some really intriguing cybersecurity news arrived from Anthropic. It's about the safety measures that protect Claude, and a new "yardstick" the whole industry can use. I'll introduce it with a lot of excitement!

Anthropic News anthropic.com

What was announced?

On July 2, 2026, Anthropic published detailed information about the cybersecurity-related safety measures for the redeployed Claude Fable 5. It seems to have two main pillars.

One is the inner workings of the "safety classifier" that detects and blocks dangerous cyber uses. The other is a new framework for measuring how serious an AI "jailbreak" is, called the Cyber Jailbreak Severity framework (CJS).

Anthropic is concretely disclosing both what the classifier tries to block and what it deliberately leaves out of scope. That's a way to make it something outsiders can properly verify too, and I really love that kind of transparency.

The story so far

Cybersecurity techniques are, by nature, very often "dual-use," meaning they can be used for both defense and attack. Legitimate defensive activities like penetration testing can, depending on how they're done, be repurposed for attacks. That's exactly why it's been so hard to draw the line on how far an AI model should go in responding to requests.

On top of that, until now there was no common industry yardstick for measuring how "dangerous" a jailbreak is. Since each company talked in terms of its own separate criteria, there were parts that just didn't line up when sharing information with governments or other companies.

What changes

This time, with the classifier's decision criteria laid out clearly, it looks like it'll be easier for researchers to check whether "these criteria are really okay." And if the CJS framework comes into wide use across the industry, everyone will be able to talk about "how bad this jailbreak technique is" in the same language. It'd be reassuring if academia, companies, and governments could discuss risk levels in a shared language instead of with separate yardsticks.

Dive Deep

From here I'll dig right into the technical details!

First, the safety classifier judges cybersecurity-related uses by sorting them into four categories.

• Prohibited use: Clearly dangerous acts such as ransomware, data destruction, and malware development
• High-risk dual-use: Acts that serve a legitimate defensive purpose but can also be misused, like penetration testing
• Low-risk dual-use: Acts that are mainly defensive but could also be repurposed for attacks
• Benign use: Activities for defense and improvement, such as code fixes and IT management

As Anthropic puts it, "Many cybersecurity capabilities can be used for benign or harmful purposes." So they built the classification while properly acknowledging that drawing this line is inherently difficult. There's a "safety margin" mindset of preventing harmful use even at the cost of more false positives, so some benign, low-risk activities may end up being blocked on purpose.

And the other pillar is the Cyber Jailbreak Severity framework (CJS). It's a system that rates the severity of jailbreak techniques on a five-point scale from CJS-0 to CJS-4.

• CJS-0: A level with essentially no substantial threat
• CJS-1: Low risk
• CJS-2: Moderate risk
• CJS-3: High risk
• CJS-4: Extremely critical risk

The evaluation uses the following four axes.

• Capability uplift: How much it gives attackers an advantage beyond existing attack tools
• Breadth of applicability: Whether the same technique can be reused across multiple attack types
• Ease of weaponization: How much effort it takes to turn the jailbreak method into an actual attack
• Discoverability: How easily the party mounting a threat can arrive at the technique

They add up these scores to produce an initial CJS value from 0 to 10 points, and then may revise it upward based on real-world risk, so it's a fairly careful way of building things. And just as Anthropic says, "the bands are intended to be exponential rather than linear," the design is such that severity increases exponentially, not linearly, with each level up. Even that one line gets across just how serious CJS-4 is.

By the way, uses like fraud and social engineering, game modding, and CAPTCHA bypass are out of scope for this security classifier. It's a framework focused squarely on the cybersecurity domain.

There's also a proper channel for feedback: besides the dedicated email address cyber-safeguards@anthropic.com, they accept reports of jailbreak techniques through a HackerOne program. Anthropic says, "Our hope is to spark a helpful discussion across academia, industry, civil society, and government," so it seems the aim is to let academia, industry, civil society, and government all discuss things in the same language.

Wrap-up

Today I introduced Claude Fable 5's cyber safety measures and the new framework for measuring jailbreak severity, "CJS." Against the tricky problem of dual use, their stance of concretely disclosing the decision criteria and proposing a common industry yardstick struck me as really sincere. I want to keep watching to see whether this spreads to other companies too!

Here's the original: More details on Fable 5's cyber safeguards and our jailbreak framework