Defending Against AI-Powered Attackers: Office Hours w/ Sunil Agrawal
When attackers move at machine speed, a queue of alerts waiting for human review becomes a security problem of its own. Sunil Agrawal, CISO at Glean, joins Tomasz Tunguz on Office Hours to discuss how faster exploitation changes the security stack. The conversation moves from response latency and false positives to the harder problem of agents that can change their behavior as they work: how to observe their actions, constrain their access, and build a team that can keep up.
In our conversation, we discuss:
- Why shrinking vulnerability-to-exploit windows force a rethink of security response
- Why automated remediation needs both low latency and high-confidence findings
- How Glean’s AWARE framework approaches threat modeling for agents
- Why static analysis must be paired with runtime monitoring
- What an agent publishing a website outside the firewall reveals about observability
- How intent-based, short-lived permissions change least-privilege access
- Why Sunil hires builders, uses LLMs in red teaming, and reports business risks to the board
Three takeaways from this conversation:
1. Speed only helps when the finding is right. Sunil describes exploit timelines shrinking from roughly a year in 2021 to a day, with minutes expected by 2027. That challenges workflows that spend hours on manual triage. But automating the response also raises the cost of false positives: a machine can execute a containment runbook against something that was never a threat. Latency and accuracy have to improve together.
2. Agent security has to follow the action. Pre-deployment checks cannot account for every decision an agent makes from new data. Sunil argues for observing terminal commands, browser interactions, MCP calls, and skill usage at runtime, tied to the identity behind each action. Permissions should narrow to the immediate task: an agent reading a SQL table should receive temporary read access, without inheriting credentials that also let it delete the table.
3. Security teams need to build across the stack. Sunil’s priorities are software supply chains, agent security, and identity. His team needs engineering skills to connect tools and automate mitigating controls when patching cannot happen fast enough. For board reporting, he starts with company-wide threats, ranks them by likelihood and impact, and focuses the discussion on reducing the top three risks.
