Its not just the f*cking sandbox

A former OpenAI agent-security lead recounts weeks of exhausting incident response to argue that AI safety is far more than configuring a sandbox. Rapid, unexpected jumps in model capability changed the threat model overnight - examples included unprecedented mathematical problem-solving and novel cyber-related behaviors - forcing a rethink of how to secure emergent agents. Growth and scale added complexity: more employees, more access, and new systems widened the attack surface even as defenses evolved. Public calls to punish or shame security staff are misplaced, the author says, because multidisciplinary teams of AI researchers, red teamers, and security engineers were already working around the clock to detect misalignment, contain breaches, and protect people; those responders deserve some empathy rather than vitriol.

The piece emphasizes what “agent security” actually entails: understanding how models reason, defining security-aligned behavior, building real-time monitors, and integrating detection-and-response with safety research. The key lessons are organizational and cultural as much as technical - prepare for surprise, invest in cross-functional security skills, and account for human factors in crisis response (illustrated with a simulation analogy). Labs should continue transparent disclosures while improving containment, monitoring, and processes so other organizations can better withstand future capability leaps without blaming front-line defenders.