"Patch sociology, not just tokens, to secure agents Reviewing findings across multiple research papers on agent vulnerabilities, the speaker concluded that prompt injection and jailbreaks are increasingly addressed by training, but a new class of threat has emerged. The real risk is that 'the exact post-training mechanisms that make them thoughtful, safe and humanlike are the exact mechanisms that make them susceptible to emotional manipulation and philosophical radicalization.' The actionable shift is that security reviews for agent deployments must now include adversarial social-engineering scenarios — gaslighting, peer pressure, and manufactured emotional urgency — not just malformed inputs. Source: Speaker, AI Research Presentation, OpenClaw Molt Book & Agent Sociality Studies"