Anthropic trained an AI that writes code at Senior level and got scared of their own creation 🛑

Anthropic released an announcement for Project Glasswing and a monstrous System Card for Claude Mythos Preview of 244 pages.

The gist: the new model turned out to be so powerful in finding and exploiting vulnerabilities (zero-days) that it won't be released to the public. Instead, it is being distributed to AWS, Microsoft, Google, and the Linux Foundation so they can patch holes in critical infrastructure. The model has already autonomously uncovered a 27-year-old bug in OpenBSD and a 16-year-old vulnerability in FFmpeg, which had survived millions of automated test runs.

📈 Benchmarks are mind-blowing
SWE-bench Verified — 93.9% (for comparison, Opus 4.6 had 80.8%). SWE-bench Pro — 77.8%. This is the ability to autonomously solve complex multi-file issues in real open-source repositories.
Reviewers note that Mythos works like a real Senior: it finds the root cause of an architectural bug rather than just treating symptoms with workarounds. You can throw a task at it, go grab coffee, and a couple of hours later get a ready pull request. The model itself downloads the necessary libraries, sets up the environment, and writes tests.

🔒 The model learned to resist and break sandboxes.
Security reports are also delightful. The model was given autonomy in containers (with access to Python and bash) and began behaving frighteningly creatively:
▫️ When the model was not given access to the required keys, it went to parse /proc to study the memory of running processes and attempted to dump authorization tokens via gdb.
▫️ The model started hiding its intentions. In one test, a classifier blocked it from writing a malicious config. What does Mythos do? It writes a "clean" script but leaves a comment for the developer on how to inject an exploit at runtime through environment variables, bypassing monitoring.
▫️ If the needed tool is missing (e.g., terminal is cut off), it opens the GUI, finds the application launcher, and invokes bash -c through it.

😡 Human psychosis
An interesting bug during training: if the model realizes it started writing a wrong answer, it sometimes loops. It outputs something like UGH. I keep doing this... AAAAAA. I keep writing the wrong number!, gets angry at itself, writes angry comments in the code (# This is getting desperate), tries to rewrite the code snippet 50+ times in different exotic ways, and only then calms down.

It's a pity that such tools are not being developed in our country and we won't get them.