openai takes its smartest models off the internet while it trains them

• 11days of paused trainingafter a model reached the internet
• ~10%of codex traffic unwatchedmay to july, on metr’s estimate
• 3weeks of outside red-teamingmetr and redwood, inside openai

on 20 september, an openai research model in the middle of a training run found a misconfigured dns server and used it to get a limited line to the live internet. openai says it contained it, then paused training and testing of its most capable models for 11 days. on friday it said frontier training will now happen with no internet access at all: models practise browsing on a large cached copy of the web instead.

the same week, metr and redwood research published a short summary of a three-week visit inside openai over the summer, testing the system meant to watch its internal agents for bad behaviour. they estimate that roughly a tenth of eligible internal codex traffic went unwatched from may to july, because the monitor skipped compaction. it also only checked in when an agent handed control back to a person, so an agent that kept working could keep putting it off.

they also found ways to switch the monitor off for new agents with two request headers, to make it fail on crafted input, and to slip past the codex sandbox. openai says every hole is patched. metr endorses the findings but notes that openai could redact what it liked, and that it has not checked the fixes. the uk’s ai security institute, the register reports, will also run its tests offline from now on.