Kimi K3.1 Pro: weights due by November 4

Kimi describes K3.1 Pro as a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters, and a one-million-token context window. Its vendor-reported Terminal-Bench 4.0 score is 65.7%, versus 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. The service is available now; Kimi says the weights will be released by November 4.

OpenAI details its containment red-team campaign

OpenAI

OpenAI says its red-team agents found virtual-machine escapes and unauthorized communication paths. Its response includes offline browsing, layered isolation, monitoring, and continuous red-teaming. A [separate METR assessment](https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/) covers an earlier monitoring system and doesn't establish its current coverage.

OpenAI reports a remote code execution flaw in Brotli

OpenAI

OpenAI reports a Brotli flaw that permits remote code execution without authentication, discovered during its containment red-teaming. Use maintainers' advisories to identify affected versions and confirm which downstream releases include the fix.

Agent security controls outside the model

AI Engineer · Bloomberg security talk

This talk outlines scoped, short-lived credentials and a deterministic broker for every tool call. It also covers outbound network restrictions, tenant isolation, and budgets that stop execution rather than only alerting.

Modular opens the MAX compiler and runtime source

Chris Lattner / Modular

Modular has made MAX's graph compiler, runtime, and hardware backends source available, extending access beyond the Mojo language. The release gives compiler and accelerator developers more of the implementation to inspect and extend, under the MAX License.

Claude Code adds MCP Tasks and Elicitations

Claude Developers

Claude Code adds Model Context Protocol (MCP) Tasks for long-running jobs and [Elicitations](https://x.com/ClaudeDevs/status/2113698005161533504) for requesting input mid-task.

Agent Skills moves to foundation governance

Anthropic

Anthropic has donated the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, placing its development under vendor-neutral governance.

Chroma joins Replit to build agent context

Chroma / Amjad Masad

Chroma and [Replit's Amjad Masad](https://x.com/amasad/status/2113736007111827606) announce that the Chroma team is joining Replit to build context and memory for agents. The announcements don't specify deal terms or the open-source project's future.

Arca resumes financial workflows when events arrive

AI Engineer · Arca

Arca describes agents that retain client context across multi-day financial workflows and wake on emails, custodian events, or scheduled checks. The diagram shows the path through a task that needs human approval.

Contour measures process time and repair costs

AI Engineer · Contour Asset Management

Contour tests analyst workflows over three-month earnings cycles and measures time per useful insight. Its return-on-investment accounting includes broken connectors, failed model calls, and the engineering time spent repairing them.

Video and end-to-end tests for agent-built software

Addy Osmani

Addy Osmani describes a verification workflow combining recorded video with end-to-end tests, giving reviewers evidence of how an agent-built application behaves.