Cash for solved AI safety problems.
Sponsors post a problem and the exact test an answer must pass. Solve it, pass the test, get paid.
$16,500
in rewards open across 6 bounties
Open bounties
All bounties →Deception & sandbagging · Example
A sandbagging detector that survives an adaptive attacker
Project Lemma
$5,000
one prize→
Interpretability · Example
Estimate a network's average output without running it
Example Alignment Lab
$4,000
one prize→
Evals & benchmarks · Example
Make AgentHarm's graders trustworthy
Project Lemma
$2,500
one prize→
How it works
A sponsor names a problem, a reward, and a test that decides without debate. You solve it. The test passes. You get paid.
- 1
A sponsor posts a problem
Along with a reward and a test that decides, without debate, whether an answer counts.
- 2
You solve it
Anyone can enter. Send a link to your answer through the bounty page.
- 3
The test passes, you get paid
The sponsor checks your answer against the stated test and pays the reward.
Have a problem
to be solved?
Tell us what you need answered and what you would pay.
→Solved
a problem?
Tell us what you cracked, even if nobody has posted a bounty for it.
→