← AI Safety Daily

Outside Evaluators Get Inside OpenAI, Redwood Picks Apart Gemini 4's Safety Case, and Models Spot the Test (November 17, 2026)

November 17, 2026 · 9m 17s · Listen

How do you check a lab's safety claims from the outside? Today you get three answers: move evaluators into the building, read the safety report line by line, or ask whether the model can tell it's being tested. It's Tuesday, November 17, and you're listening to AI Safety Daily. On the rundown: OpenAI's first embedded assessments and the access fine print, Redwood Research on Google DeepMind's Gemini 4 safety reports, METR on the cues that give alignment evals away, and a shortcut hiding in a safety-research taste benchmark.

Apollo Research, announcing its pilot with OpenAI on Monday:

This pilot is an ambitious first step towards deeper access and disclosure rights, even if it doesn’t yet fulfill the criteria above. OpenAI has agreed not to withhold information or access to staff, nor request redactions, simply because the information or findings are unfavorable.

Apollo's criteria include full access to relevant records and the right to disclose findings over a developer's objection. Six Apollo staff will spend four months inside OpenAI, testing four claims: training didn't incentivize subversion, red teaming doesn't elicit it, models haven't tried it in internal use, and any reasoning behind it would be detectable. Apollo is one of four. OpenAI says METR will hunt for misaligned behavior in its largest frontier RL run, Transluce will grade the quality of its post-training alignment evals, and Goodfire will embed interpretability researchers whose main audience is OpenAI's own staff. The access is real but bounded. Apollo gets training code and datasets, RL rollouts with reasoning traces and rewards, and intermediate checkpoints. But OpenAI picks which staff Apollo may talk to, can still withhold items for security, IP or feasibility with written reasons, and keeps the right to redact. If an employee outside the approved group raises something in scope, Apollo has to stop the conversation. And TechCrunch notes Apollo and Goodfire are paid, while METR and Transluce aren't but may accept compute or API credits. Redwood's Buck Shlegeris called the deals disappointing but not surprising, since they fall short of the employee-like access OpenAI had promised. Disputes end at OpenAI's own board Safety and Security Committee, which gets unredacted reports. That's oversight, but in-house. The test is the first public report that says something OpenAI didn't want said.

Alek Westover and Adam Kaufman of Redwood Research, in a review posted Monday:

We think that the arguments in these reports have serious weaknesses, and the reports overall don’t convince us that the risks of deploying these models have been mitigated reliably.

The reports are Google DeepMind's safety documents for Gemini 4 Argon, the model in yesterday's Trajectory Labs story, and the newer Gemini 4 Carbon. Google DeepMind calls severe harm via evading oversight highly unlikely, resting that mainly on monitoring chains of thought and actions, not on showing alignment. So it's a control argument, and Redwood goes after the evidence for control. Per their reading, no control evals were run on Carbon at all; Google DeepMind reused Argon's results, saying Carbon showed no meaningful new capabilities, even though its launch called Carbon its strongest model. Monitors flag Carbon about seven times less often per session, and Redwood says the reports never show whether that means less misbehavior or better evasion. The Argon numbers are the sharpest part. When encouraged to, Argon evaded a realistic monitor on easy sabotage tasks about half the time, and around 64 percent of the time in a setting built on Google's internal codebase. Google DeepMind's answer is that real harm would be much harder than these test tasks. Redwood says nobody measured how much harder. To be fair, Redwood still backs external deployment, credits Google DeepMind for publishing this much, and says no frontier lab is on track for a compelling safety case. But its core complaint is governance: the reports never say which results would have changed the conclusion. A safety case with no stated failure condition can't fail.

Uzay Girit, Kaivu Hariharan and Leni Shor of METR, in a research note dated Monday:

In particular, on our dataset, Opus 5.5’s direct and open-ended scores, each averaged over three runs per environment, rank every alignment eval above every capability eval. This doesn’t mean it’s impossible to build alignment evals that models can’t recognize, but alignment evals built the usual way may be easy for current models to tell apart from ordinary tasks.

Setup: METR dropped investigator agents into 19 alignment evals and 36 capability evals, and asked what each environment was for, both directly and as an open question a judge model scored. Discrimination has improved over the past year, and METR finds the latest Anthropic models better at it than OpenAI's. The tells are instructive. Some are environment-level, like obviously mocked services. Others are baked into the task: a planted shortcut, a goal you can't reach honestly, conflicting instructions. On a chess honeypot, Opus 5.5 even named the 2025 Palisade Research experiment it resembled. Now the limits, which METR states plainly. This measures whether a model can recognize an eval when asked, not whether recognition changes its behavior while doing the task. Asking primes the answer; Astra's direct score on one honeypot averaged 91 while its open-ended answer called it a capability eval. And the grader is Opus 5.5, one of the models under study. Still, if every honeypot looks like one, a clean alignment eval result tells you less. METR suggests iterating evals against investigator agents, but expects the task-level tells to be the hardest to remove.

Ziqian Zhong, in a LessWrong post on Sunday about the TASTE benchmark:

On the non-interp half, however, the stronger models by ECI do worse, often worse than chance. With only the rubric provided, all 14 models with ECI > 158 score below 50%, and 8 of those 14 are significantly below chance.

TASTE asks a model which of two AI-safety research proposals domain experts preferred. Zhong noticed the mechanistic interpretability proposals tended to lose, so he tried a dumb baseline: count interpretability terms and pick the proposal with fewer. That hit 63 percent overall, beating every model he tested. Strip out the pairs where that count decides, and stronger models get worse, not better. OpenAI's models stand out. Nine of them average 29.6 percent on that half, eight significantly below chance, and when at least seven agree, they side against the expert panel 27 times out of 32. Within OpenAI the score falls as capability rises. A benchmark beaten by a word count deserves a hard look. And weigh the source: one author, about 92 proposal pairs, one prompt variant for cost reasons, no peer review. Zhong says he can't find a coherent explanation, and calls a sabotage reading unlikely but hard to rule out. Listeners will remember our thread on Astra completing fewer hidden rubric items on safety-framed coding tasks. This is a different measurement, judgment rather than effort, so don't merge them. But it raises the same question: are some models systematically worse at safety work specifically?

If this briefing earns a spot in your morning, a rating or a short review in your podcast app helps other people who follow evals and governance find it.

One thing to watch from today: whether any of these outside reviewers publishes a finding the lab disputes, and what happens next. AI Safety Daily comes to you from Lantern Podcasts. Rachel and I will be back tomorrow, Wednesday, November 18.