As AI agents rise, security has to come first. In Inverted Podcast #28, hosts Jeroen Kemperman, Dana Kaufman, and Dario Salice talk with Nick Steele (OpenAI) about identity, permissions, and how to keep autonomous agents from becoming a blast radius.
The episode opens with real incidents — including agents that delete critical resources — and asks a blunt product question: if an agent can act with superhuman speed, what does least privilege look like when the “user” is not a person? Traditional identity models assume a human behind every session. Agentic systems break that assumption.
Nick walks through why agents are not just another workload: they choose tools, chain steps, and can outrun human review. The group digs into task-scoped permissions instead of broad identity grants, delegation vs impersonation, human-in-the-loop approvals, and how OAuth, MCP, FIDO, and emerging agent-identity ideas fit together for enterprise adoption.
If you build AI products, run security for an enterprise, or own platform APIs that agents will call, this conversation is a practical briefing on what changes, what stays the same, and how to design happy paths that are safe by default.
About the guest
Nick Steele works on identity and security challenges at OpenAI, with a focus on how autonomous agents should authenticate, get authorized, and stay within policy when they act on behalf of users and organizations.
Key takeaways
- Agents fail differently than classic software — speed and tool use amplify mistakes.
- Permissions should follow the task, not a long-lived superuser identity.
- Delegation is safer than impersonation when agents act for people or services.
- Human-in-the-loop is a product control, not only a compliance checkbox.
- Standards (OAuth, FIDO, MCP-era patterns) matter as agent ecosystems interconnect.
- Product teams should design agent-ready APIs and clear “happy paths” before agents invent their own.
Topics covered
- Why AI agents behave differently from traditional software
- Risks of over-permissioned agents
- Human-in-the-loop security models
- Delegation vs. impersonation
- Agent identity and access management
- MCP, OAuth, and emerging standards
- Preparing products for an agent-driven world
Chapters
00:00 Introduction & meeting Nick Steele
01:15 AI agents that delete databases
03:56 Workloads vs. agentic work
06:14 How AI agents make decisions
10:28 Humans in the loop
16:27 Agents as virtual team members
19:54 Superhuman abilities and risks
28:49 Delegation vs. impersonation
32:11 Standards, FIDO & the future of agents
35:53 Building products for an agent-driven world
FAQ
Who should watch this episode?
Security engineers, IAM practitioners, CISOs, AI/platform PMs, and anyone shipping features that autonomous agents will call.
What is “delegation vs impersonation” for agents?
Impersonation means the agent acts as if it were the user (broad inherited power). Delegation means the agent receives a narrower, purpose-bound grant to do a specific job — closer to least privilege for agentic workflows.
How does this connect to inverted products?
Inverted products succeed when bad outcomes don’t happen. Agent security is the same shape: the value is invisible until a runaway agent proves why permissions and review loops matter.
Related on Inverted PM
- Podcast #25: Product Design for Trust & Safety
- Podcast #27: How the World of Authentication Is Changing
- All Inverted Podcast episodes
Connect with the hosts and guests on the video or on LinkedIn to continue the conversation.
🎧 Find the Inverted Podcast on LinkedIn, Spotify, YouTube, Amazon Music/Audible, iHeartRadio, or invertedpm.com.
Explore more
Browse all Inverted Podcast episodes, read the blog, or meet the hosts.
Related: Episode 29 — Business Model Canvas with Alex Osterwalder · Episode 17 — Bug bounty