Won asks companies to name failure thresholds that halt model deployment

Thank you so much.

This is for all panelists.

What is the exact failure rate or benchmark threshold that would cause you to halt the deployment of a new model?

And if your model successfully generates a bioweapon design 5% of the time during internal testing, do you still release it?

And then I have a separate question.

Will you commit today to providing exact redacted logs of the prompts that successfully jailbroke your models by biosecurity filters during your internal testing?

So to your question on statistics or probabilities, I don't have a specific number.

But what I can say is that no level of catastrophic risk is remotely acceptable.

And we don't believe that we should be training models unless we can make an extremely strong case that we can keep them under human control.

Council Member, I would just note that we are committed to building AI safely, and we have laid out in detail how we approach this in our scaling framework and in our preparedness reports.

In relation to further detail, I would need to get back to you on specifics.

For the other panelists?

Yes, ma'am.

I don't have a specific number for failure rate that would cause us to make a non-launch decision, although I do know that we aim for safety and performance standards ahead of allowing ourselves to launch.

And if we're not meeting that, it's not a good enough model or product for us to release.

I'm not personally aware of any...

Bio security evaluations where jail successful jailbreaks happened.

So I'm going to have to go back to the teams and ask if they've ever seen such a thing and get back to you on that.

Thanks.

And as I mentioned earlier, I think we were essentially the first team in the industry to design and test models for this purpose.

We outline a very nuanced process of how we decide how to launch with respect to biological capabilities, for example.

And in fact, in the past, we have held a model because we needed more time to assess its biological capabilities and proactively implemented safer safeguards for this reason.

To the point of jailbreaks, this is a great question.

One of the hard parts is sometimes you do or don't want to share a jailbreak depending on whether it would put other models at risk.

And so we regularly publish our research and alignment reports with all the information that we possibly can that would be safe to do so.

And we try to set the industry standard there.

Right now, I know you're not in the room, but behind me, we have the safety claims and we have all the incidents that have happened, including biosecurity, which you have heard also from Councilmember Lynn Shulman, who continues to lead us as a committee chair for health and hospitals.

So can you please clarify?

So Anthropic has been the only one that gave me a clear answer.

So for Anthropic, what is the exact failure rate and benchmark threshold?

And the rest of you have not given me that answer.

And then we'll conclude and move on to Council Member Hanif.

The answer is it's nuanced.

What we do, if you would like the sort of complex answer, is we detail multi-hundred pages of reports, several dozen of which are usually allocated to biology every time we launch.

You can see the exact levels of capability and the exact reasoning that we put in when we look at a number of different tests that have different, so on, you know, quote unquote, failure rates that allow us to make that decision.

I think that highlights just how much effort we put very specifically into not just biology, but cybersecurity and self-improvement and alignment as well.

And everybody else?

Because if what you claim is true, then you wouldn't have released these models that are causing these harms.

But I will press on.

We have to move on.

Thank you so much.