Two outside groups measured how much of OpenAI's agent traffic was eligible for monitoring
METR published on 22 October an account of a three-week assessment of OpenAI's internal misalignment monitoring, run jointly with Redwood Research from 21 July to 11 August 2026. One employee from each organisation worked on OpenAI-issued devices that stayed on the premises, with access to OpenAI's internal codebase, the monitor's prompts and configuration, and internal documentation. METR estimated that roughly 75 to 94 per cent of agentic internal Responses API traffic from models at least as capable as GPT-5.5 was eligible for monitoring, and 99.2 per cent of the subset used by internal Codex agents. METR says the figures describe the system as it stood on and before 21 July.
The account lists routes around the monitor. Monitoring fired when an agent returned control to the user, so while compaction events went unmonitored, an agent that kept acting could “indefinitely postpone being monitored”. OpenAI staff had found that gap before the exercise and closed it during it. The report ran to 61 pages; OpenAI redacted it to 38 for the two organisations' leadership. Buyers of agentic systems can ask a vendor what share of agent traffic is covered and who measured it. OpenAI says all the vulnerabilities the two groups identified that affect its production safeguards stack have been patched. METR writes: “We have not evaluated these claims.”
Cadence turns an upgraded Munich data centre into a sovereign AI platform for chip design
Cadence said on 21 October that an upgraded Munich data centre now gives European customers a sovereign platform for designing chips, packaging, systems and AI infrastructure “while maintaining data residency within European jurisdiction”. Cadence OnCloud on sovereign infrastructure is “available today to European customers” and is certified to ISO 27001, ISO 27017 and SOC 2 Type II. Nvidia Vera and Grace CPU systems, the Cadence Millennium M2000 supercomputer and Reality Digital Twin Platform instances are planned as part of the ongoing Munich expansion rather than live. Cadence calls itself the first EDA vendor to deliver a comprehensive sovereign AI platform for semiconductor, system and AI factory design in Europe.
The three partner companies quoted in the release are there for the Reality Digital Twin Platform, not the Munich facility: Vertiv designs, simulates and validates its modular MegaMod HDX AI data centre with it, HPE is adopting it for its Data Center AI Mod POD, and France-based Nation Data Center uses it to design and validate sovereign, high-density AI infrastructure. A European design team can ask Cadence which workloads run on the sovereign infrastructure available today and which wait for the planned capacity. The release names no customer for the Munich data centre.
Quick Hits
- Microsoft is adding EU-partner oversight of engineer remote access. — Data Guardian and the tamper-evident Production Touch Ledger are rolling out to EU and EFTA customers, covering qualifying Microsoft engineer remote access to in-scope production systems (Microsoft, 20 October).
- HeyGen Voice Turbo tops Artificial Analysis's controlled-voice leaderboard, listed at $4 per million characters. — It scores 1202 Elo, in a 1-3 rank band shared with HeyGen Voice (1195 Elo, $30) and Qwen-Audio-3.1-TTS-Plus (Artificial Analysis).
- Nvidia's DSX POD is a reference design for AI factories partners build inside modular containers. — Nvidia says deployment at a prepared site drops from six months to as little as 14 days, with worldwide delivery planned in early 2027 (Nvidia).
- Mistral and Nvidia say they will publish public benchmarks and open datasets for physics models and engineering agents. — The 21 October collaboration pairs Mistral's physics AI with Nvidia's CUDA-X and Warp libraries and Omniverse (The Decoder).