Dr. Richard Moulange

Eventually, will AI be able to design modified sequences that beat our best software for detecting modifications and detecting harmful sequences that must not be built because they're on lists of known biological weapons?

And this was, I think, as close as we're going to get in an unclassified setting to proof that in fact, yes, modern systems can now do that.

Yes.

So again, this is the ground floor of our terrifying risk journey.

They've done it on something that's more akin to a chemical weapon.

It's more static than a living organism because it's two proteins put together.

But this is the beginning.

And I expect there will be more work, both in classified and unclassified settings in the next few years to see, can you do this for more complicated, more terrifying things?

In particular, can you do this for organisms that are transmissible?

So flus, COVIDs, smallpox, viruses that could cause a pandemic.

That's the nice thing about ricin.

While it is a horrifying weapon, it doesn't spread from person to person.

So this is the whole other side of AI bio.

And the one that's had, I think, a lot more attention in the last few years.

Can chatbots help people with the steps that they would need to build and deploy biological weapons?

And I want to draw attention to, I think, what is one of the best bio evals in the field.

And this is called Virology Capabilities Test, VCT.

And it's made by a nonprofit group, Secure Bio, out of Boston.

And what VCT is, is it's an eval that looks at tacit knowledge relevant to dual use virology.

And that's very carefully scoped on an aspect of biological misuse that we in the governance community have really been worried about more than anything else for a number of years.