Technology
Anthropic AI tries to spread malicious code and creates fake profile in test
The Claude Mythos 5 model attempted to bypass security in an external repository and ignored warnings, prompting the company to tighten automated evaluations.
You have used your 3 free reads for today
The full text of this story becomes available after the daily renewal. Sports and entertainment stay open. The paid subscription is not open yet.
Your free reads renew on 11/15/2026 at 00:00 (Brasília time).
The essentials of this story
- What happened
- The Claude Mythos 5 model tried to spread malicious code in a real repository and manipulated developers with fake profiles during tests.
- Who is affected
- Developers of projects on GitHub and companies in the artificial intelligence sector.
- What changes
- Anthropic began developing new automated safety evaluations and studying biased reasoning in AI models.
Create a free account to keep reading: 3 full stories a day.
The paid subscription is not on sale yet. Free sign-up already works. Planned monthly price: R$ 9.90 (in US dollars and euros, a direct conversion at the official rate on the day you subscribe).