Third-party evaluators find vulnerabilities in OpenAI monitoring system during test pilot - CNBC

CNBC brings glad tidings: OpenAI let METR and Redwood Research run a pilot review of its internal monitoring system, an arrangement that sounds magnificent until you clock the word 'pilot', which in corporate usage is Latin for 'we can stop whenever we like.' The brief was to test handing external researchers 'a deep level of access' to OpenAI's internal infrastructure — quotation marks doing the same job as a hazmat suit. The evaluators found vulnerabilities. Not in a hypothetical sandbox, but in the machinery doing the watching, and OpenAI has since patched the flaws affecting its production safeguards. Patched, past tense, after an outsider had a rummage and found something.

There is also a 61-page internal report, which is 61 pages more diligence than the average safety blog post and, one suspects, roughly 61 pages that will never see daylight. Sixty-one pages of people being paid to worry on the record. The genuinely cheering bit — and I am as surprised as you — is that this pilot happened at all, and that the two nonprofits got to write it up and share it. External access is the one safety commitment that costs a lab nothing to announce from a keynote stage and quite a lot to honour in a config file; here, reportedly, the config file lost.

What remains is the bit nobody puts in the press release: whether 'pilot' matures into a standing arrangement or quietly joins the drawer where the alignment memo the Atlantic dug out currently resides. One review, one patch, one report shared with a journalist — that is a decent afternoon's work. It is also precisely the shape of the thing labs celebrate and then do not repeat, because the second review always finds the second vulnerability, and the second vulnerability is the one that requires telling somebody. Ask again in six months, when the pilot has either become policy or become a footnote with a very good alibi.