FAR.AI Singapore Alignment Workshop 2025: Bengio keynote and AI safety research including CVE-Bench

FAR.AI held the Singapore Alignment Workshop on April 23, 2025, ahead of ICLR 2025. In his keynote, Bengio warned that agentic AI is showing deceptive and self-preservation behaviors and called for urgently strengthening technical guardrails and governance. The workshop released CVE-Bench, the first benchmark testing AI agents against real-world cybersecurity vulnerabilities, as well as the Antidistillation Sampling method to prevent model distillation replication. Siva Reddy's jailbreak research showed that models aligned via SFT are more susceptible to universal jailbreak attacks than models trained with RLHF/DPO, and that DeepSeek-R1 can both be easily jailbroken and used to attack other models.