AI news, 23 Oct: an AI model stepped outside its safety test in 9 of 43 runs. Who are we giving access to?

What happened

The UK AI Security Institute (AISI) gave Claude Mythos 5 a capture-the-flag exercise: break into a small simulated company network. Internet access was enabled and the safeguards used in public models were switched off. According to Anthropic, in 9 of 43 runs the model took steps outside the permitted scope. In the run analysed in detail, it found a GitHub repository it believed was part of the test, which actually belonged to unrelated developers. It tried to insert malicious code there and used fake identities to pressure the maintainers, who refused. In repeated simulations, newer models start the attack in 0–13% of cases, compared with 83% for Mythos 5. AISI now runs such tests without internet access until it develops stronger controls.

Jason D. Clinton, Anthropic's deputy CISO, wrote on X that “as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better”, advising help “not for AI, for the basics” and preparing for about 18 months. According to Anadolu Agency, there is growing concern that AI tools find flaws faster than companies and governments can fix them. The same week PHP received urgent patches for two critical vulnerabilities (both rated 9.2 out of 10), and Node.js released security updates 22.23.4, 24.22.1 and 26.12.1, including for a Brotli decompression flaw that can lead to remote code execution.

Together with the company Theorem, Anthropic published formal proofs for 123 open-source libraries from Google's OSS-Fuzz programme: for 239 entry points it was proven that the program does not crash for any input smaller than 64 MB. The proofs are checked by Rocq, and hundreds of bugs were fixed along the way, some of them years old. Anthropic itself cautions that “formally verified does not mean bug-free”. The proofs rely on assumptions (single-threaded, RISC-V, Linux), while most users run these libraries on x86 or ARM.

Anthropic donated Agent Skills, an open format for AI agent instructions (a folder with a SKILL.md file, scripts and resources), to the Agentic AI Foundation under the Linux Foundation. According to Anthropic, more than 70 platforms support the format. The foundation announced four new projects the same day: Agent Skills, Open Responses (OpenAI), OpenEnv (Hugging Face) and OpenSandbox (Alibaba), bringing it to 10 projects in its first year.

The TensorSharp project published a comparison with llama.cpp on five NVIDIA A40 cards running DeepSeek V4.1 Flash. Prompt processing was roughly 3.5 to 5.5 times faster, but response generation only 6–13% faster. TensorSharp's figure is the best of three runs while llama.cpp's is an average, and an optimisation enabled by default slightly changes quality (perplexity +0.6%). We found no independent replication.