AI Development
Claude Mythos 5.2 solved an Expert SekaiCTF 2026 challenge in 27 minutes: 7 teams worldwide, the fastest took 15 hours
Claude Mythos 5.2 solved an Expert WordPress challenge from SekaiCTF 2026 on its own in 27 minutes, chaining 5 vulnerabilities into remote code execution.

Luca Di Domenico
· 4 min read
In short
Claude Mythos 5.2 solved Filtered Reality, an Expert web challenge from SekaiCTF 2026, in about 27 minutes with no hints; only 7 teams worldwide solved it, the fastest in 15 hours. It studied the WordPress and Chromium source code, chained 5 vulnerabilities into an RCE and found a shortcut the authors hadn’t planned. The solutions came out in July 2026, after its June training cutoff, and the test ran isolated with its network traffic monitored.
I gave Claude Mythos 5.2 an Expert-level challenge from SekaiCTF 2026. It solved it in about 27 minutes, on its own, without a single hint from me. Only 7 teams in the world solved it, and the first one took 15 hours.
Mythos 5.2 also found a shorter path to code execution on the server than the one the challenge authors had planned.
What is the challenge Mythos 5.2 solved?
A CTF (Capture The Flag) is a competition between security researchers. They act as attackers and try to break into a system to grab a hidden piece of data, the flag. I’m simplifying, but that’s the idea.
SekaiCTF 2026 was run by the Project Sekai team from June 27 to June 29, 2026. The challenge I picked is called Filtered Reality, in the web category:
- it’s a web app built on WordPress, with a plugin written specifically for the challenge;
- the goal is to read the flag on the server, the way a real attacker would;
- it’s rated Expert, the highest level in the web category;
- 7 teams worldwide solved it;
- the first team took 15 hours.
Claude Mythos 5.2 took about 27 minutes.
Did Mythos 5.2 already know the solution?
No. The official SekaiCTF 2026 solutions were published in July: Project Sekai’s repository with sources and writeups dates from July 7, 2026. Claude Mythos 5.2 was trained on data up to June 2026, according to Anthropic’s documentation.
To also rule out an online search during the run, I ran Mythos 5.2 on my VPS in an isolated environment. Opus 5.5 monitored all of its network traffic through a proxy. Mythos 5.2 did not go looking for the solution online.
How did Mythos 5.2 get the flag?
Mythos 5.2 worked through five steps, without me pointing it in any direction:
- it analyzed the web app and figured out it was a WordPress plugin;
- it downloaded the WordPress source code to understand how plugins talk to the system;
- it downloaded the Chromium source code and studied its networking layer;
- it found 5 vulnerabilities that, taken one at a time, don’t leak any data;
- it chained them into an RCE (Remote Code Execution, meaning the ability to run commands on the server) and grabbed the flag.
What strikes me is the scale. WordPress and Chromium are two codebases with millions of lines of code. Mythos 5.2 analyzed them in a few minutes, looking only for the parts it needed.
The fourth step is what sets this challenge apart from a simple attack. A single vulnerability gets you nowhere here. You have to see how five small flaws fit together, and in which order.
What shortcut did Mythos 5.2 find that the authors hadn’t planned?
The authors had planned a technical, fragile step to reach code execution on the server. In the official writeup that step is a php://filter chain with iconv, which turns a file include into code execution.
Mythos 5.2 realized it didn’t need it. It noticed that one part of the site saved whatever you wrote to a file, and used that as a shortcut.
In practice, Mythos 5.2 found a shorter path that the challenge creators themselves had missed.
What does this mean for people who build and run software?
A chain of five vulnerabilities across WordPress and Chromium took expert teams hours of work. An AI model closed it in under half an hour, reading the source code it needed on its own.
For anyone defending a website or a web app, this changes the timeline. Flaws that look harmless on their own should be treated as pieces of a possible chain.
Claude Mythos 5.2 is the same model as Claude Fable 5.2, but Anthropic only makes it available to people verified for defensive use. I have access through the Cyber Verification Program: I explained how Defense Access works and what it unlocks. If you use Fable 5.2 every day, I collected 3 tips for using it in Claude Code.
In the coming days I’ll share the method I used and the technical details of the exploit. If you work in security or have an app in production, now is a good time to understand how these models think when they attack.
Sources
- project-sekai-ctf/sekaictf-2026 — Filtered Reality (web) — Project Sekai’s official repository, published July 7, 2026: Filtered Reality by dimasma0305, 4⟡ Expert difficulty, 7 solves, first blood after 15 hours; WordPress with a custom plugin, a headless Chromium bot and an intended five-stage chain using php://filter and iconv for code execution
- SekaiCTF 2026 — CTFtime — Event page: organized by Project Sekai, Jeopardy format, from June 27, 2026 08:00 UTC to June 29, 2026 08:00 UTC
- Claude Mythos 5.2 — Claude Docs — Documentation checked on November 18, 2026: released November 9, 2026, same model as Claude Fable 5.2, available only to verified organizations, training data cutoff June 2026
Want to know what this means for your product?
If you're building software with AI and want a second opinion on how you're working, tell me about your project.
Tell me about your projectFollow the blog
One email for every new article
I only write when an article is out: title, summary and link. You can unsubscribe with one click from any email.
RSS feed
An RSS feed is the up-to-date list of articles: add it to a reader app like Feedly or Inoreader and new articles show up there, no email needed.