Anthropic releases Claude Fable 5 with Mythos limits
The safety story is a routing decision, and the compliance story is 30 days of retained frontier-model traffic.
TL;DR
CyberScoop reports that Anthropic is publicly releasing Claude Fable 5, the same underlying model as Claude Mythos, while routing some cybersecurity and biology queries to Claude Opus 4.8. Anthropic says internal and external red teams found no known universal jailbreaks, but did not specify whether partial jailbreaks appeared. Fable and Mythos traffic will be retained for 30 days across Anthropic and third-party platforms, aligning with the White House’s voluntary frontier-model access framework.
Anthropic is trying to make Claude Mythos public without making all of Mythos public. According to CyberScoop, Claude Fable 5 uses the same underlying model as Mythos, but queries in narrow cybersecurity and biology areas are routed to Claude Opus 4.8, a previously public model, rather than answered with the full Mythos capability. That matters because Anthropic had previously treated Mythos as too capable to release broadly, especially for hacking and bioweapons research.
The operational claim is narrower than the marketing frame. Anthropic says Fable 5 went through internal and external red-team testing and that testers found no known universal jailbreaks. CyberScoop notes the company did not say whether partial jailbreaks were found. For security teams, that distinction is not academic: older AI models have repeatedly been jailbroken, and Anthropic itself expects adversaries with financial motives to try to bypass the safeguards.
The compliance hook is the retention change. Anthropic will keep all Fable and Mythos user traffic for 30 days on its own platforms and third-party services, and says the data will not train new Claude models or be used for non-safety purposes. That 30-day window tracks the Trump administration’s scaled-back AI executive order, which created a voluntary framework for AI companies to give the federal government pre-release access to frontier models for up to 30 days, according to CyberScoop’s earlier reporting at https://cyberscoop.com/donald-trump-white-house-ai-executive-order-scaled-back/.
For practitioners, Fable 5 does not create a new federal mandate by itself. It does change the risk conversation around frontier-model use: retained traffic, third-party platform coverage, safety classifiers that may block benign security work, and an explicit admission that the model’s uplift would be valuable to cybercriminals. That is enough to put Fable and Mythos into AI use policies before the first enthusiastic engineer pastes vulnerability data into a prompt.
Published ·Deep Fathom