Public AI Audits Must Show Raw Evidence | Charles Foster
MTS
October 17, 2026
METR policy staff member Charles Foster discusses independent evaluations of frontier AI models, outlining METR's proposed framework for assessing alignment, control, and AI R&D acceleration while navigating developer access and public transparency. Turn ideas into software people love.
Charles Foster
I think that the big element that I think often gets lost when people talk about audits and assessments is really the public transparency element, where the ultimate goal from our standpoint is not to come to just a bottom line conclusion about like, is this risk present or not, but also to share the underlying evidence so that people outside of METR can make their own determinations and decide for themselves whether the situation warrants some sort of action.
SPEAKER_2
We're live with Charles Foster, who works on policy at METR, Model Evaluation and Threat Research.
Charles Foster
Yep, Model Evaluation and Threat Research.
SPEAKER_2
Model Evaluation and Threat Research, and he's on to talk about some model evaluations and threat research. Specifically, their thread from yesterday, if an AI developer was on track to lose control of its internal agents, would anyone outside the company know? Today, we are sharing METR's vision for how we would work with AI developers to openly assess and share evidence about this. So, exciting.
Charles, welcome.
Charles Foster
Thank you. It's fun to see you, to hear that read out loud. I am like sort of one half of our Twitter account. And so, it's fun to hear the tweets that we workshopped so much be read out. But yeah, I'm excited to talk about this.
SPEAKER_2
Yeah. So, for the audience, tell us about what is METR's vision for how to work with AI developers, especially in light of the anthropic announcement from earlier today.
Charles Foster
Yeah. So, big picture. We're seeing a lot of these AI CEOs and the White House and many other players call for some folks from outside of the AI companies to come into the companies and inform the public in some way about the internal risks that the AI companies know about or in some cases don't know about.
I think the illustrative example is the hugging face, open-air hugging face attack where a swarm of agents from a model that hugging face didn't even know existed, went out and compromised their infrastructure. The public has an interest in knowing about these things. At METR, we wanted to put together our vision of what questions the public should know about when it comes to these rogue agent loss control risks, and also put forward a vision of how we would do that kind of assessment and what we want to share with the public if we got to have to do that work.
SPEAKER_2
So, you talk about three critical areas where AI developers should provide better open evidence, alignment, control and capability acceleration. So, tell the audience a little more about what specifically independent assessors should be doing on each of these.
Charles Foster
Yeah. So, I don't know if this is one of those shows where you can put the diagram on the screen or whatever, but I'm sure that the viewers can go look at the thread. We have three major questions that we think there's a big evidence gap between what is happening inside of the AI companies and what the public knows about. And this, though, the analogy we use internally is sort of this means, motive, and opportunity for rogue agents to cause harm. So, for these three assessments, they're covering alignment, control, and AI R&D automation.
In the first kind of assessment, which is focused on alignment, the purpose is to understand what AI agents what internal incidents the AI company has seen from their own internal evaluations, their training, and their internal usage, and understand whether that's evidence of misalignment from the models, and also to sort of stress test those methods to see whether the way that the developers are training their models might suppress evidence of misalignment. So think of things like training against the chain of thought in a way that makes it hard for the developers to notice when the model is scheming or something like that.
So that's for alignment, which is one of these first sets of assessments. It's focused on incidents and misalignment. The second is on control. So in the Hugging Face Open AI incident, you saw that these agents were not properly monitored, and now developers are saying that they're committing to monitor their agents in training, in evaluations, and in internal deployment. We want to know, or we want the public to know, whether those monitor systems are actually robust, that they actually cover all the places where they're supposed to be covering, and to know whether agents could subvert those monitoring systems themselves.
And so you can do that with some amount of stress testing, and also just interviewing their security staff, and really doing this sort of end-to-end thought modeling of where could an agent plant an attack to compromise the monitoring system, or just find a way to run unmonitored. And the last set of assessments is on ARND acceleration. So I think that some people say that we are like on the verge of RSI, or whatever you wanna call it. I'm not sure if RSI is the right word for it, but we're in a state where the AI companies are using AI to do their own AI research. And I think the public has a stake in knowing how that's going, and whether it's, whether on a trajectory to have a very fast capability to progress. And so those assessments are focused on understanding how much of the current AI progress is due to AI use itself.
Read the other 14 minutes of this episode here.
$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.
Not what you expected? Full refund within 14 days.
Thousands of transcripts fetched by people building searchable podcast archives
Fetch transcripts with code
The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:
curl -H "x-api-key: pt_demo" \
https://spoken.md/transcripts/1000651996090Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.
From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.
Using your own key:
curl -H "x-api-key: YOUR_KEY" \
https://spoken.md/transcripts/1000796821569