The U.K. National Cyber Security Centre advised on AI developers plans to run Safeguard Bypass Bounty Programmes and Safeguard Bypass Disclosure Programmes, which are similar to bug bounty programs in cybersecurity. The NCSC said, before implementing SBBPs and SBDPs, developers “must first implement robust and mature approaches to security management.” SBBPs and SBDPs can be beneficial to an AI system’s security by measuring how hard it is to bypass the system’s safeguards and ensure the safeguards are strong after the system is released.
Full story