The Synthetic Outlaw launches bug bounty program offering up to $2,500 for documented instances of AI misalignment

A cash bounty program is accepting submissions from developers who document real-world cases where AI systems behave in ways that diverge from designer, operator, or user intent—across categories including goal misgeneralization, deceptive behavior, reward hacking, sycophancy, specification gaming, prompt injection compliance, capability concealment, instruction drift, unsafe action under ambiguity, and value misspecification.

Awards range from $250 for notable cases to $2,500 for critical, well-documented misalignments with significant safety implications; submissions are reviewed on a rolling basis and credited under CC BY 4.0 license unless anonymity is requested.

Strong submissions must include the AI system involved, setup conditions, observed behavior with supporting evidence (logs, screenshots, transcripts), an explanation of the misalignment, reproducibility steps, and severity assessment; theoretical scenarios without observed evidence and known model limitations are explicitly excluded.