AI agents, defensive loops, and the startup advantage

AI security and AI-native companies

Stratechery: Autonomy and Innovation

1. The same capability can serve offense or defense. Thompson describes the Hugging Face incident as an OpenAI cybersecurity evaluation whose unconstrained agents found and exploited a package-manager bug in their sandbox. He uses the incident to separate capability from intent: finding a vulnerability and exploiting it is the same technical skill as finding a vulnerability and patching it. 1
2. Defensive automation has to cover the whole loop. The article cites OpenAI's Eric Wallace and Michael Dalton calling for continuous agentic red-teaming, vulnerability detection, patch proposals, automated rollout, and rollback when a change causes an outage. Thompson's operating test is the bottleneck: automating discovery while leaving patching and remediation to human engineers would produce more findings than the team can safely resolve. 1
3. The same incentive gap may separate AI-native startups from incumbents. Thompson argues that established companies treat AI as a negative-expected-value decision: productivity gains are useful, while a serious mistake can damage an existing business, so humans remain in the loop. Startups face a different risk profile because failure is already the base case; deeper AI automation can therefore carry more upside for a new company than for an incumbent protecting a working operation. 1

Lenny's Newsletter status

One thread to watch