This week, AI labs gated their cyber models and rebuilt the plumbing under coding agents

Between November 1 and November 7, OpenAI, Anthropic and Google released no new general-purpose text flagship. Their announcements covered the layers around the models instead: who gets the most capable cyber models and under what controls, the infrastructure coding agents run on, and benchmarks that measure where agents still fail. One major open-weight model also arrived when its maker said it would.

Security work moved from pledges to gated products and pipelines

The biggest release of the week was GPT-6 Cyber on November 5. It is a model for finding, reproducing and fixing vulnerabilities, and OpenAI rates it Critical for cybersecurity under its Preparedness Framework, the same rating as GPT-6 Astra. Access is narrow by design. Organizations need Daybreak Red approval plus SSO, MFA, managed devices and IP allowlisting, and individuals need an invitation, identity and liveness checks and a FIDO2 hardware key. The API model page lists $20 per million input tokens and $100 per million output tokens, with a 1.05-million-token context window. In OpenAI's own honeypot tests, the model went out of scope 0.07% of the time without deployment safety instructions and 0% with them. No independent evaluation is available yet.

OpenAI published more security work around that launch. According to its news feed, on November 4 it described an autonomous pipeline that found, validated and generated patches for 948 local privilege-escalation bugs in the Linux kernel. Earlier in the week it raised its bug-bounty payouts, and on November 6 it announced a partnership with Hacktron to probe its own systems. We could open only OpenAI's feed summaries, not the full posts, so we can't say how many of those kernel patches have been accepted upstream.

Anthropic's security news was about controls for customers. On November 5 it added a security center to Claude Enterprise that compares an organization's settings with Anthropic's recommended baseline. It also made inference hooks generally available: these send prompts and pending tool calls to a company's data-loss-prevention tools. A beta lets an organization's own endpoint detection and response (EDR) sensor run, in detection-only mode, inside the local Cowork virtual machine. Both labs are acting on the same assumption. Agents that can act on systems need oversight inside the loop, not just a policy document.

Coding agents got new infrastructure, and two acquisitions

On November 2, Anthropic put message threads into public beta on the Claude API. After the first request, a client sends only the new messages plus the previous response ID. Token usage and pricing stay the same. Anthropic reports a drop of about 35% in median time to first token on 500k to 1M-token requests, and a thread expires after five idle minutes. The same day, OpenAI described Pando, a Git-compatible backend it built to handle the volume of code Codex generates.

Two acquisitions followed on November 4. OpenAI said it plans to acquire Pierre Computer Company to expand Codex's tools for building and reviewing code. Meta confirmed to TechCrunch that it bought Conductor, an app that runs several coding agents in parallel on other labs' models; Conductor's cloud service shuts down on November 20. Hugging Face also proposed small CPU-only integration contracts for Transformers models. In a replay of five months of commits, one of these contracts flagged an attention bug that had sat unnoticed in main for almost two months.

New evaluations measured where agents still fail

Several benchmarks this week tested agents on how reliable they are rather than how impressive their best results look. Ai2's StateBench found that frontier models lose track of variables that change over long inputs: accuracy fell to 13% or lower on the longest, hardest problems. Ai2's ToyBlocks found that an agent writing its own tool descriptions after testing the tools raised task success to 63%, compared with 36% for descriptions written by a human expert. OpenAI added LifeVal, for everyday tasks, and BioBasicsBench, 100 tasks on raw wet-lab data. For anyone building agents, StateBench suggests verifying state explicitly instead of trusting a long context window, and ToyBlocks suggests that tool descriptions are worth iterating on empirically.

Open weights: one promise kept, one still pending

Moonshot's Kimi K3.1 Pro weights are now public on Hugging Face, as the company promised for this week. The model card lists a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters and a context of about 1 million tokens. It ships under a custom Kimi license, not a standard open-source one, and its benchmark figures are Moonshot's own. On November 4, Tencent said Hy4's weights and system cards would follow "in the coming days." When we checked on Saturday evening, Tencent's Hugging Face account still showed only the August Hy4 preview.

Infrastructure kept running into local politics

On November 5, Google announced a $17 billion data center in Wyoming. It plans to power the site partly with a 702 MW natural-gas fuel cell plant with carbon capture, built with AEP. In the same week, OpenAI launched a Grid Protectors program with utilities and published a compact for building AI infrastructure that protects local communities. Both companies now present power supply and community impact as part of the case for building, not as details to settle later.

The throughline

Read together, this week's announcements show the labs competing on what surrounds the model. That includes gated access for dangerous capabilities, controls that sit inside the agent loop, faster plumbing for long sessions, and benchmarks that expose specific failure modes. For developers, the practical changes were lower latency on long Claude sessions, a new and expensive gated cyber model, and growing evidence that long-context state tracking and tool descriptions need deliberate engineering.