Anthropic Mythos isn’t as scary as it sounds (yet)

Thousands of zero-days, exploits across every major OS, “AI hacker” narratives everywhere.

But when you look at the data, the story is more grounded.

Anthropic ran large-scale autonomous campaigns, thousands of parallel agents, that cost roughly $20k–$25k to surface a handful of meaningful bugs.

Impressive, but it also tells you where the boundary is.

This is still a probabilistic search problem.

Models like Claude Opus 4.6 already succeed at full exploit generation a meaningful amount of the time. Mythos improves that mostly by running more attempts in parallel: generate candidates, test them, discard failures, keep the rare wins.

That is persistence at scale, not a fundamentally new capability.

So what changed?

Not what’s possible, but the cost and time curve.

Bug discovery is stochastic. Success per run is low but non-zero. Compute scales linearly. Given enough trials, sufficiently complex systems will yield bugs. Mythos just compresses the timeline.

Further, there’s also a big difference between finding a bug and finding something operationally useful. Many outputs do not generalize, require chaining, or break under slight environmental changes. Even then, someone still has to validate what matters.

Which leads to the real takeaway

Cost is the bottleneck, not capability.

A $20k–$25k campaign to find a single high-impact bug is not trivial. Scale that across targets, and economics still dominate feasibility.

Mythos matters. But for now, it changes the slope of the curve more than the nature of the game.

‍