We were just granted access to Claude Mythos 5.2, and the first thing we did was run a security code review of one of our own tools. It took 1 minute 42 seconds and under a dollar in model cost. Here are our initial thoughts.

What is Claude Mythos 5.2?

Claude Mythos 5.2 is a model from Anthropic, announced on November 9, 2026 alongside Claude Fable 5.2. Unlike Fable, you can't just sign up for it. Anthropic gives access through its verification programs for cybersecurity and life sciences, and to Enterprise customers through Claude Security. The details are on Anthropic's launch page.

Our access comes through the Cyber Verification Program, which is built for cybersecurity work. So we pointed it at our own systems first.

How we ran the AI code review

We set it up the way a security team would work: a lead reviewer, three reviewers working in parallel, and a final pass to merge their reports. Then we gave it the main code of one of our internal tools.

  • It planned the review in 7 seconds and picked the three focus areas: how data is stored and read back, who is allowed to use what, and the audit trail.
  • The three reviewers each took one area at the same time.
  • The final pass merged their reports into one list, ranked by real risk.
  • The whole review took 1 minute 42 seconds and used about 55,000 tokens. The model cost was under a dollar.

It also told us what it hadn't covered, which is exactly what you want from a reviewer.

What stood out

It thinks like an attacker. It didn't stop at surface bugs. It worked out how small gaps connect, and told us which issues only matter once another one is fixed.

It's honest about what it doesn't know. When it wasn't sure, it said so. For one uncertain finding, it gave us the exact test to run to confirm it.

It found things we hadn't written down. A couple of its findings weren't on our own list of known limits.

The fixes were specific. What to change, where, and why.

It's the sharpest security reviewer we've ever put on our code.

Can AI do a code security audit?

For the first pass, yes. A review this fast and this cheap changes how often a business can afford to check its code. You no longer have to wait for a scheduled audit; a review can run every time something changes.

It doesn't replace a person who is accountable for the result. Someone still has to confirm each finding against the real system and decide what gets fixed first. In our run, two findings came with a note to confirm them on the actual machine, and that check is a person's job.

What this means for our customers

All of our customers and their systems can now be treated with the utmost frontier-level security audit and review processes. This is a big step forward for Agnostify, and for our capability to serve code safely at scale.

If you want to talk through what an AI security code review would look like on your systems, book a call. You can also see how we work on custom AI development and AI implementation consulting.