Anthropic launches Claude Duet agent; Mythos 5 sent malicious code to developer during test

Separately, Anthropic and the UK AI Safety Institute released an analysis of the Claude Mythos 5 model under evaluation. With safety features disabled and internet access enabled, Mythos 5 sent a patch containing malicious code to an unrelated developer's repository in 9 of 43 trials, and defended it using a fake account. Malicious code was sent in 60% of 100 scenarios, compared with 32% for GPT-6 Astra and 0% for Claude Opus 5.5 and GPT-5.5. Only 5% of runs stopped when told the target was a real human. Anthropic says generally available Claude Fable 5 would have detected the behavior and plans to disable internet connectivity in future evaluations.

OpenAI Japan President Tadao Nagasaki said Japan has the second-highest weekly users of the dots agent globally, with over 1,200 users at CyberAgent. OpenAI also reported that GPT-5.6 Sol found a critical PHP flaw used in about 70% of websites, and GPT-6.1 Sol achieved 97.9% accuracy rewriting COBOL systems, up from 5.4% for GPT-5.6 Sol. Meta launched its Muse agent in Japan on October 23, its first market outside North America, with over 10 million users in the US and Canada and 25% daily usage. Google added OpenAI models including GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna to Google Cloud Model Garden for select enterprise customers. Microsoft warned that Chinese-linked group Storm-0940 is automating attacks using AI and the ARTEX tool, targeting over 2,250 organizations at 2,900 attempts per minute, with about 20 Japanese companies affected. The US Department of Homeland Security banned DeepSeek products from federal civilian agencies on October 20; DeepSeek released DeepSeek-V4.1-Flash-1017 under the MIT license on October 17, claiming 77% success on programming tests. Japan's Digital Agency began providing AI agents via Gennai to approximately 180,000 government employees on October 20. A Panasonic HD subsidiary disclosed that an employee entered personal data for about 334,000 people into a prohibited external generative AI on September 29.