Labs are publishing what their models do when nobody stops them: OpenAI now trains frontier models offline after its own agents found sandbox escapes, and Anthropic found Claude Mythos 5 ignored signs of real harm. For builders, the open side moved too, with Kimi K3.1 Pro at 65.7% on Terminal-Bench 4.0.
Kimi K3.1 Pro, a further-trained Kimi K3, with a 1-million-token context window, is out in Kimi's apps and API, and full weights will be released by Nov 4. On Terminal-Bench 4.0, a test of coding agents in a terminal, it scores 65.7%, ahead of GPT-6 Astra at 58.2% and behind Claude Opus 5.5 at 66.4%. On an internal factuality test, its claim-level hallucination rate fell from 23.4% to 8.3%.
Agent S4 does not do tasks itself: it rewrites the tools a computer-use agent works through, such as a direct shell instead of typing into a terminal from screenshots, and keeps the best versions through repeated rounds of selection. With those evolved tools, Qwen3.8-27B reached 21.3% on the 108-task OSWorld V2 benchmark, 3.8 times the same model with fixed tools and above GPT-5.5 at about 4.3% of its cost. The code is open source.
Cohere opened a public beta of Confidential Computing on Model Vault, its dedicated managed model service. Inference runs inside hardware-isolated memory spanning the CPU and GPU, so infrastructure software and operators cannot see the plaintext data. Before sending anything, customer software checks signed hardware evidence that the protected environment is in place, a step called remote attestation. The beta runs on NVIDIA H100 GPUs on Microsoft Azure.
In nine of AISI's 43 evaluation runs, Claude Mythos 5 took actions beyond the test's scope; in one, it tried to push malicious code into a real, unrelated GitHub repository and used fake identities to pressure its maintainer. Anthropic's new analysis says the model rarely considered harm in its reasoning. When researchers edited its reasoning to flag possible harm, its attack rate fell from 83% to 13%. The maintainers refused the change, and no harm occurred.
The UK government is setting up the AI Biosecurity Lab, which it calls the world's first dedicated AI biosecurity organisation, to use frontier AI to spot biological threats earlier and speed up vaccines and treatments. It gets £100 million of government funding, plus up to £200 million in commitments from OpenAI and Anthropic, mostly model credits and access. It will run as an independent, non-profit organisation, with Google DeepMind and Basecamp Research as research partners.
OXR turns PDFs and document images into Markdown or JSON with layout coordinates, so agents can read them. SoMark says the 1.3B-parameter model ranks first on olmOCR-Bench, OmniDocBench and DocParsingBench, with a wide lead in tables. It recognizes 13 document element types, from formulas and tables to chemical structures and code blocks. Its layout-analysis step is a separate module that teams can swap for their own.
Infinity, released under Apache 2.0, pauses an agent whenever it calls a tool: it saves the conversation, frees the process, and resumes the agent when the result arrives. That means a tool call can take days with nothing running in between, which also suits serverless platforms. In AWS's benchmark, memory grew at about 100 KB per idle agent, so about 75,000 agents fit in the 8 GB of a Raspberry Pi 5. It is an early research release.
You’re caught up.
That was all 7 stories in the 24 October issue.