Skip to main content
MANIFOLD
Will a top 10 AI lab be victim of an attempted data poisoning attack by EOY 2027?
1
Ṁ100Ṁ10
2027
55%
chance

Resolves YES if an AI lab which ranks among the top 10 AI labs by some reasonable model quality metric reports an attempted (not necessarily successful) data poisoning attack against their models. Also counts if some sufficiently credible third party (such as an embedded evaluator) gave such a report.

Research on data poisoning gives the impression of it being really easy to do, and it being able to implant an undetectable backdoor into a model through well-hidden phantom signatures in a small amount of data. But if that's true then we'd expect to actually see a data poisoning attack soon.

"Writing for the LLMs" doesn't count and probably wouldn't be reported on anyway.

Get
Ṁ1,000
to start trading!