Securing the Agentic Enterprise: Office Hours w/ Jonathan Jaffe
AI gives attackers new capabilities, but defenders have access to the same models. Jonathan Jaffe, CISO of Lemonade, joins Tomasz Tunguz on Office Hours to discuss what that means for enterprise security. From agents that investigate vulnerable dependencies to policies that govern individual actions, Jonathan describes how security teams can use engineering to keep pace. The starting point is understanding the business’s tolerance for risk and building controls around the work people and agents actually do.
In our conversation, we discuss:
- Why stronger models can help defenders close vulnerabilities faster
- How to align AI experiments with the business’s appetite for risk
- Why everyone on Lemonade’s security team is an engineer
- How internal agents connect threat intelligence to code investigations
- Why each agent needs an identity and policies that follow its actions
- How package proxies, pinned dependencies, and adoption delays reduce supply-chain exposure
- Where AI accelerates security investigations and incident decisions
Three takeaways from this conversation:
1. AI can shorten the defender’s response loop. More generated code creates more opportunities for bugs, but Jonathan also sees models improving code review, testing, and vulnerability remediation. Lemonade’s internal security agents connect threat intelligence with repository analysis, including checking whether a vulnerable package’s affected method is actually used. That context helps the team distinguish an urgent exposure from a finding that does not create the same risk.
2. Agent permissions need context at the moment of action. An agent’s identity is only the beginning. Jonathan argues that policy decisions should account for who initiated a task and what the agent is trying to accomplish when it acts. As organizations deploy more specialized agents, static access grants become less sufficient. Security teams need a way to register agents and connect their actions to the intentions and permissions behind them.
3. Security is an engineering and business discipline. Jonathan’s team builds software to investigate and reduce risk. Its choices also depend on what the business is willing to accept: a monitored, time-bounded experiment can justify different controls from a critical production workflow. Practical measures such as pinning dependencies and delaying new packages complement that approach. The goal is to make deliberate decisions about exposure while giving the company room to learn.
