Anthropic details Fable 5's cyber safeguards and proposes a severity scale for jailbreaks

On July 2, 2026, Anthropic described in detail how the safety classifiers in Fable 5 work and released a first draft of an industry framework for rating the severity of jailbreaks.

How Fable 5's safety classifiers work

Cybersecurity is inherently dual-use: the same capabilities can serve defense or attack. That is why Anthropic does not try to block everything related to it. Instead, Fable 5's classifiers rate each request against four categories:

• Prohibited use: actions that in most cases can cause significant harm and have no defensive value. Blocked unconditionally.
• High-risk dual use: tools widely used by attackers but that also have legitimate applications. Also blocked.
• Low-risk dual use: capabilities that are mainly defensive in nature but could theoretically help attackers too. Monitored; sometimes blocked under the safety margin.
• Harmless use: legitimate tasks with no potential for harm. Allowed, with monitoring.

The key concept in the system is the "safety margin": a deliberately widened zone in which the classifier blocks requests out of caution, even when they look potentially harmless. A request has to look clearly safe to be guaranteed to pass. For Fable 5, Anthropic deliberately made this margin wider than in its previous models.

Why does the industry need a single jailbreak severity scale?

A jailbreak is an unconventional way to make an AI model bypass its own restrictions. It can be nearly harmless (removing just one minor restriction) or critically dangerous (unlocking a broad range of harmful capabilities and making the model fundamentally more dangerous). The industry has no common terminology for assessing these risks, which seriously hampers dialogue between AI companies and regulators.

Anthropic, together with its Glasswing partners, has prepared a first draft of such a framework. The company sees it as a starting point for a broad discussion involving academia, industry, civil society and government bodies. Proposals are accepted at [email protected].

"We believe that by working together, we can develop a standard that allows this technology to be used for defensive purposes while preventing abuse,"

Anthropic said in its official statement.

Anthropic has also launched a program on the HackerOne platform: security researchers can officially submit jailbreaks they find in Fable 5 for review by Anthropic's team.

What this means

Anthropic openly acknowledges the fundamental difficulty of cybersecurity as a dual-use field and chooses a calibrated approach over blanket blocking. The jailbreak severity framework is the first serious industry attempt to create a shared language for AI labs and regulators. If the standard catches on, companies will be able to describe safety-bypass threats in agreed terms, and governments will be able to assess the risks of new models more precisely.

Frequently asked questions

What is the "safety margin" in Fable 5?

The "safety margin" is a zone in which the classifier blocks requests out of caution, even if they do not look clearly harmful. This lowers the risk of accidentally allowing dangerous tasks, at the cost of some false positives. In Fable 5, Anthropic deliberately widened this zone compared with previous models.

How can researchers report a jailbreak they find in Fable 5?

Anthropic has launched a dedicated program on HackerOne: security researchers can submit jailbreaks they discover for official review. Conceptual proposals for the severity framework are accepted at [email protected].