Critique of current AI safety bug bounty programs (opens in new tab)

The potential value of AI safety bug bounty programsGenerally, AI labs should (and most do) put their models under extensive safety testing before deploying them to prevent misuse, scheming, and other dangerous behaviors. This may include internal tests, red-teaming efforts by third-parties, etc. However, edge case safety vulnerabilities will likely slip through, and these can still cause damage. If any of the risks of AI systems from labs that implement strong safety measures in the first pl...