Critique of current AI safety bug bounty programs
AI safety bug bounty programs, while valuable for identifying post-deployment vulnerabilities, are currently too narrow in scope and ambition. OpenAI's program, for instance, focuses on material harm and requires high reproducibility, with low payouts and few rewarded issues since its July 2025 inception. A more effective approach would broaden the scope to include a wider range of risks and lower the reproducibility threshold to capture critical vulnerabilities that may not be consistently replicable. AI
IMPACT Critiques of current AI safety bug bounty programs highlight the need for broader scope and lower reproducibility thresholds to better capture critical vulnerabilities.