After an accidental data leak revealed the existence of a new AI model called "Mythos," Anthropic has confirmed it's testing a system representing what the company calls a "step change" in capabilities. The confirmation comes at a particularly awkward time — right as Anthropic fights a legal battle with the Pentagon over defense contracts.
The leak itself was embarrassing for a company that markets itself on AI safety and responsible development. Internal documentation about Mythos appeared briefly on a public cloud storage bucket before being quickly removed. But the internet never forgets, and screenshots circulated widely enough that Anthropic had no choice but to comment.
What does "step change" actually mean?
In AI, "step change" is the kind of phrase that can mean everything or nothing. It's vaguer than "breakthrough" but implies something more significant than "incremental improvement." Based on Anthropic's statement, Mythos represents advances in reasoning capability, multi-step problem solving, and what the company describes as "more robust alignment."
Translation: it's better at complex tasks and supposedly harder to jailbreak. Whether that constitutes a genuine step change or is just positioning for marketing purposes remains to be seen.
I've watched enough AI hype cycles to know that "step change" often means "we improved our benchmark scores." What matters is whether Mythos can do things that Claude 3.5 genuinely couldn't — and whether those things are actually useful rather than just impressive in demos.
Anthropichasn't released technical details, benchmarks, or examples of Mythos in action. That's either because the model is still early in testing, or because the reality doesn't quite match the hype. Given the leak forced their hand on acknowledgment, I'm betting it's the former.
The Pentagon problem
The timing of this leak is particularly interesting because Anthropic is currently fighting to access to Pentagon contracts — a position that seems at odds with the company's founding mythology about responsible AI development.
