The reviewer agreed with the conclusion and doubted the method
A software team that had to learn again on 800,000 devices, eight compromised certificates patched in a day, and METR's outside review of Anthropic's risk report.
How we Organise
A team that knew software had to learn the work again on the devices
Kuldeep Singh’s team modernised a retail sensor fleet of about 800,000 battery-powered devices across 3,000 stores. The team was strong in software and thin in embedded firmware. Its own write-up says documentation was not enough. They built a device lab, measured power and radio behaviour on real hardware, and earned thirteen new working defaults. Each default came from a problem they hit.
This is our talent position stated back to us with the credentials already in the room. The team did not lack experience, and it still had to demonstrate the capability on the live system before the work could move. The account is one team’s, on a practitioner site, and it measures nothing against the alternative.
Capability is demonstrated on live work or it is not demonstrated.
How we Build
Eight compromised certificates were patched out of one network in about a day
On 19 October the private keys of eight widely trusted eMudhra intermediate certificates were found compromised. A forged certificate for any domain became possible. Cloudflare had its clients, tunnel and CDN patched by the next evening. Browsers moved fastest, because their trust stores already take rapid updates. The lesson the engineers state is short: a trust store cannot be a static list.
That is the posture we ask clients to take on everything they run. The static list of trusted certificates was the protection for years, and on one day in October it was the exposure. The open question is how many firms can update their trust stores at that speed, and this is one responder’s own account.
Currency is safer than stability.
How we Assure
The outside reviewer agreed with Anthropic’s low-risk call and not with the method
METR reviewed the misalignment section of Anthropic’s August 2026 risk report, under a veto Anthropic never used. It agrees the risk from the covered models is low. Its reason is that their capabilities are limited, not that the safeguards hold. It is not confident the report’s method would identify the risk if it rose, and it recommends red-teaming Anthropic’s monitoring end to end.
Our risk position asks for exactly this check. This is also the second source in ten days: on 12 October we covered Gans and Holden’s audit model, which found independence alone does not deter an agent that can learn the audit schedule.
Every model this team runs is validated before it is used and after every material change, by somebody independent of the people who built and run it, as deeply and as often as the firm’s model risk policy requires for its tier.
Together the two say independent validation is necessary and not sufficient. METR’s red team is the kind of draw the audited system cannot predict.