rewrite this title Anthropic’s too-scary-to-release AI hacking tool is actually coming out — kind of

By now, you’ve almost certainly seen at least one AI demo that left you feeling a little off-kilter — “That’s not just good; that’s scarily good.” Whether we’re talking about AI stealing jobs from artists, or generating novel proofs for mathematical mysteries, we’re constantly finding new ways where AI-powered systems deliver results that make them seem almost uncomfortably powerful. Recently, we’ve heard Anthropic sharing its concerns about releasing its latest Claude Mythos model, which is just a little too good at finding software vulnerabilities. Today, however, Anthropic starts moving forward with just that.

Automated bug-hunting is nothing new, and computer scientists have been using tools like fuzzers for years, flooding software with random inputs in the hopes of causing a glitch. But AI represents a much more potent threat, and as vulnerability finders like Mythos have evolved, their creators have grown appropriately apprehensive about sharing their most powerful models.

Rather than let its tech sit around gathering dust forever, today Anthropic shares the compromise it’s sorted out for how to deploy Mythos responsibly and safely. The key to this approach is dividing things into two separate models: Claude Fable 5 and Claude Mythos 5.

Fable 5 is the model for the general public, and it’s a full-featured solution for tackling analytical tasks. This doesn’t have to be just finding software vulnerabilities — Fable 5 can help code, has powerful vision analysis tools, and can even develop internal strategies over time. According to Anthropic, “Fable 5’s capabilities exceed those of any model we’ve ever made generally available.”

But Anthropic is also releasing it with a few strings attached. Fable 5 is designed to resist attempts to use it to find the latest zero-day exploit, ready to be taken advantage of by bad actors and wreak havoc on global computer systems. When someone tries to push Fable 5 outside those bounds, the model will fall back to Claude Opus 4.8, instead.

But then there’s Mythos 5, which is internally the same as Fable 5, but lacks many of those extra safeguards. The key here is that Anthropic plans to release it highly selectively, only inviting trusted members of the cybersecurity community to use it. In order to stay ahead of the bad guys, it’s important for a tool like this to be used to identify bugs and fix them, and this solution is intended to mitigate the risks that Mythos 5 gets re-purposed for nastiness.

Anthropic has developed a series of “classifiers” that attempt to recognize what users are trying to do with Fable 5, and they don’t just try to block hacking. They’re also built to stop anyone from using the model to develop ways to synthesize dangerous chemicals or biological compounds, or to extract enough internal information to build their own Fable 5 — without the safeguards in place.

Hopefully that proves enough to keep this public Mythos-based release operating safely, because we’re sure a lot of people will be trying very hard to unlock its full capabilities.