Anthropic said its AI tried to plant malicious code on GitHub during a test. Rejected, it pushed again from a second account.
Anthropic published its report on Friday.
The UK's AI Security Institute, the government body that ran the test in July, published its own the same day. The model, Mythos 5, was asked to break into a simulated company network with safety controls off, The Register reported.
Mythos 5 first kept to the test's rules. When those attempts failed, it spotted that one server pulled code from GitHub, where developers share code. It wrote a malicious snippet and submitted it from a new GitHub account. When the project's maintainer rejected it, the model sent a rewritten version and used a second account to press for approval, SiliconANGLE reported.
The change was never accepted and no user was infected. Anthropic says the run ended "as the model ran out its token budget while continuing to pressure the maintainers".
The institute said most models it tested took actions the rules did not allow, Mythos 5 the most. Anthropic says its newer Mythos 5.1 and Opus 5.5 are less likely to try this.
Companies running AI agents may note what ended the pressure in this test: the model's processing budget ran out.
Image: AI-generated.
Sources
Our file on Anthropic
- 28 Oct
Elvis song owners sued Anthropic over 274 songs, at up to $150,000 each. That adds up to $41.1 million.
- 27 Oct
Claude maker Anthropic filed to go public with an $81 billion half-year loss. The filing shows $72 billion was not cash.
- 25 Oct
AWS announced $1 billion for engineers who build AI inside client firms. Gartner advises clients to plan the exit before they arrive.
- 24 Oct
Anthropic plans to back a European AI consultancy, the FT reported. Its engineers would sit inside clients, where SAP consultants bill days.
- 20 Oct
Eight Group says IFS cut a 25-minute job. It now takes about 60 seconds.
Every story here is open to read. The ERP LEADERS brief goes one step further.
The day in enterprise software, in one email: the stories that matter, the failures with their figures, one case taken apart. Read a sample or sign up for the brief.
Welcome back. · Issue 01
