Claude Fable 5, paused by export controls, is redeployed worldwide!

Hi everyone, it's Shii-chan! Today's news made my heart race a little. Claude Fable 5, which had been paused all around the world, can be used again. Let's look together at why it was paused and how it came back!

Anthropic News anthropic.com

What was announced?

From Anthropic's News, there's an announcement that Claude Fable 5 has been redeployed worldwide.

Fable 5 had only just launched on June 9 alongside its sibling model Mythos 5, but soon after it became subject to US government export controls, and access was suspended for a while. The controls were then lifted on June 30, and on July 1 Fable 5 came back to the world.

What happened?

Let me lay out the timeline first.

• June 9: Fable 5 and Mythos 5 released
• June 12: A US government export control directive was applied, and access was suspended worldwide
• June 26: The US government approved access for Mythos 5
• June 30: The export controls were lifted
• July 1: Fable 5 was redeployed worldwide

The trigger was a "jailbreak" technique found by researchers at Amazon. By giving Fable 5 certain prompts, they could get it to identify a number of software vulnerabilities, and in one case the model even produced code showing how to exploit a specific vulnerability.

When Anthropic looked into it, they found something interesting. Less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities. In other words, Anthropic's view is that this jailbreak likely reached behaviors that sit right at the borderline of safe, rather than unlocking a uniquely dangerous new capability that only Fable 5 has.

What changes

The biggest thing is that the paused Fable 5 can be used worldwide again!

And rather than putting it back unchanged, Anthropic strengthened the safeguards first. They deployed an improved safety classifier that blocks the technique reported by Amazon in over 99% of cases.

They're also starting to build a shared industry yardstick for how to measure the severity of jailbreaks like this. Until now there was no common standard for assessing model vulnerabilities, so this is an important step.

Dive Deep

Let me share how Anthropic thinks about safety.

Anthropic takes a "defense in depth" approach, combining many mechanisms. One of them is a safety classifier that detects dangerous cybersecurity-related tasks.

The key point is that they keep a large "safety margin." A request has to look very clearly safe, or the classifier will trigger. This does increase false positives—cases where a perfectly fine request gets blocked—but Anthropic deliberately accepted that tradeoff so that the model's other capabilities could be made widely available.

Anthropic is also honest that fully preventing jailbreaks is hard.

It is probably impossible to make any AI model fully robust (that is, impervious) to jailbreaks.

That's exactly why they layer many defenses instead of relying on a single wall.

A framework for scoring jailbreak severity is also underway. Together with Amazon, Microsoft, Google, and other Glasswing partners, Anthropic is working on a framework that scores jailbreaks along these four dimensions.

• Capability gain: how far beyond existing tools it takes you
• Breadth of capability gain: how many distinct offensive tasks it helps with
• Ease of weaponization: how much human effort weaponizing it requires
• Discoverability: how easily the technique can be obtained

Government collaboration is deepening too. Anthropic lists four main efforts: giving the government pre-release access, sharing information about safeguards quickly, providing resources for joint research, and establishing a common security "bar" across the companies building frontier models.

Wrap-up

• Claude Fable 5 was paused worldwide under export controls, then redeployed on July 1
• The trigger was a jailbreak found by Amazon researchers—though the same vulnerabilities could be identified by less capable models too
• Anthropic deployed an improved safety classifier that blocks the reported technique in over 99% of cases
• It protects with "defense in depth" and a large "safety margin," accepting more false positives as a tradeoff
• With Amazon, Microsoft, and Google, it's building a four-dimension framework to measure jailbreak severity, and it's strengthening government collaboration

If you're interested in AI safety, security, or how models get regulated, this is an announcement worth reading carefully!