pavel-gurevich (co-founder of Guardicore, acquired by Akamai; now founder and CEO of Tenzai) builds an AI agent that does automated penetration testing on top of frontier models. His take on Anthropic’s restricted “Mythos” model: the capability jump in offensive security is real, but limited, scarcity-based access is mostly brilliant marketing that hurts defenders. He prefers broad access with identity verification and accountability (OpenAI’s trusted-access approach) and open models. Expect a period like the early-internet worm era: many bugs found in critical systems, then cleanup. The game becomes how fast you find and fix. Advice for companies: run continuous automated security testing in the release process, raise human pentest frequency from yearly to quarterly or monthly, and test integrated staging environments, since many issues are logic and integration gaps rather than code bugs. Also: harness engineering as continuous build-and-remove, nightly eval labs costing hundreds of thousands of dollars a month, competing with your own model vendors, measured AI acceleration (10–20% overall, ~3× on new products), and why enterprise sales should go back to in-person.
Show: ignore-instructions
Concepts: ai-cybersecurity · harness-engineering · llm-evals · ai-moats · ai-rd-rollout · ai-hype