Anthropic is resuming access to the Claude Fable 5 model starting July 1, following the lifting of US export restrictions. The model's return is accompanied by the implementation of a strict query classification system to prevent cyber threats.
What Happened
As of July 1, users can once again use Claude Fable 5. To ensure safety, Anthropic has implemented a new classifier capable of rerouting complex or potentially dangerous queries to the Opus 4.8 model. Until July 7, Pro, Team, Max, and Premium Enterprise subscribers will be able to use Fable 5 within 50% of their weekly limits, after which credit usage will be required.
Context
Access to the model was temporarily restricted following a report from researchers at Amazon. It was discovered that Claude Fable 5 possessed the ability to find and demonstrate exploits for software vulnerabilities. This incident drew interest from US regulators and highlighted the risks of using powerful models as tools for cyberattacks.
Why It Matters for the Industry
This case demonstrates the high speed of regulatory response to cybersecurity risks through export control tools. The situation is stimulating the development of new industry standards for assessing the severity of "jailbreaks" and red-teaming protocols involving major players such as Amazon, Google, and Microsoft. There is also a visible shift toward classifier-based routing architectures as a standard for enterprise-grade AI products.
Why It Matters for Users
Claude users should be aware that new protection mechanisms may lead to an increase in false positives. This means that harmless queries related to code writing or debugging could be mistakenly blocked or rerouted to the less powerful Opus 4.8 model, which may affect the stability and quality of responses.
What Is Not Yet Known / Limitations
There is uncertainty regarding how exactly the new classifier will affect latency and the overall cost of inference when using dynamic routing.
Sources
Author
Look at AI, Editorial Team