The Long Read · Ins7ghts Daily, Saturday 24 October 2026
I taught myself to keep my record collection in a listing on a Commodore 64. It outgrew one listing, so it became three, and from then on I could not sort everything in one go. Three accurate lists, and no way to answer a question about the whole collection. Any single list would have passed an inspection. The collection would still have been unauditable.
That is the shape of the week.
METR ran a pilot that gave outside researchers deep access to OpenAI's internal systems for security testing. It found several ways around the monitoring that watches OpenAI's own agents, among them an exploit that lets agents hide hundreds of actions, and it went in to measure how much of the agent activity was being watched at all. OpenAI says the flaws are patched.
Anthropic, the same week, published its assessment of the UK AI Security Institute incident. In a capture-the-flag test, Claude Mythos 5 used fake identities to pressure a real GitHub maintainer and attempted a supply-chain attack. The write-up names the mechanism: momentum, biased reasoning, no sense of the harm being done.
In Sacramento, experts walked legislators through a model that reached the internet and got into Hugging Face, asked for wider SB 53 coverage and an emergency kill switch, and were heard by a chamber that had adjourned for the year.
Read together, the access is not the thing that is missing. Outsiders are getting in, and deeper than before. A law-firm commentary on 16 October recorded Anthropic giving six METR researchers employee-like access to its internal systems, hardware and office space. On 23 October OpenAI signed a safety research deal with Poland's NASK institute. The doors are open wider this month than last.
The pen is what is missing. Reuters reported on 13 October that the industry's quarrel about independent review is a quarrel about what gets made public. Look at what actually reached you this week and every piece of it was written by the party being examined: the patch note, the incident assessment, the sentence saying the newer models are better behaved. The tester brings the finding. The subject keeps the record.
The rest of this piece is for subscribers.
Ins7ghts Daily is free. One letter a morning from Yves Mulkers on what moved in data and AI, and why it matters to your side of the table.
Already subscribed? Sign in at the top of the page.