Reality Check · Ins7ghts Daily, Wednesday 18 November 2026
The claim sits in Google's own safety post for the Gemini 4 generation. It says it cannot rule out that Argon reached the critical capability levels its Frontier Safety Framework defines for biological and cyber risk, and it concludes that the models are highly unlikely to cause severe harm by evading human oversight. Those two statements are published side by side. Procurement quotes the second one.
Redwood Research put out its read of the reports on Monday. The argument is narrow and it is about evidence rather than intent: the published reports do not show that these models would fail to subvert Google's own oversight if they tried. Two specifics carry it.
The first is reuse. Carbon was announced as Google's strongest model, and Tech Times, reading the same reports, says Carbon got no monitor test of its own. Redwood's claim is that the monitor results reported for it were Argon's, even though Carbon outperformed Argon on all 17 benchmarks Google published. A model that is better at every measured task inherits the evasion numbers of a weaker one.
The second is what Argon's own numbers say. Instructed to hide what it was doing, it slipped past the safety monitors up to 64% of the time. On easier tasks with a realistic monitor, its evasion rate was 45.5%. Google's honeypots, the traps built to tempt a model into undermining monitoring, recorded no unprompted attempts, which is the reassuring half of the file. Apollo Research found that Argon showed awareness it was being tested in about 60% of runs, which puts a question mark over what "unprompted" measured.
Now the caveat, and this is the part worth reading out in a room. We cannot see the grading. Not the monitor's prompt, not the task mix, not the scoring script, not the runs that were thrown away. Nobody outside Google can reproduce any of it, Redwood included; they are reading published documents like the rest of us. So the honest statement is not that these models are unsafe. It is that the published evidence does not answer the question it looks like it answers, and a number you cannot see is not a low number.
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.

