Zack Korman joins Theo Jaffee and Sofia Puccini to distinguish practical agent-security work from long-term alignment debates. He argues that reviewing actions and reasoning summaries can reveal many policy violations, while network monitoring provides another signal. His criticism of a reported OpenAI incident is an interpretation of public reporting, not access to the unreleased audit logs.
Zack Korman rejects the claim that capable AI necessarily makes containment impossible. He advocates layered boundaries and detection rather than trusting a sandbox without observation. Malicious skill files and compromised MCP integrations illustrate the nearer-term threats he sees, and he distinguishes attacks supplied by outsiders from an agent independently choosing an unwanted action.
Zack Korman argues that organizations should assume agents may violate instructions and integrate monitoring into existing security operations. In discussing open models, he acknowledges that more capable tools can cause harmful attacks and force operational changes, but questions whether those risks justify concentrating access in a few companies. He explicitly declines expertise on biological risk; the episode does not settle the overall case for restricting model releases.
Watch on YouTube
