Identity for AI Agents: Least-Privilege Delegation, Consent, and Audit When Agents Spawn Agents

🤖 Identity for AI Agents: Least-Privilege Delegation, Consent, and Audit When Agents Spawn Agents

📅 Friday, September 11 🕗 08:00 PST / 11:00 EST / 17:00 CEST

Cloud native solved workload identity with SPIFFE, mTLS, and least-privilege service accounts. AI agents break most of those assumptions. They are created at runtime rather than deployed, they spawn and delegate to other agents, and they call tools, MCP servers, and APIs using long lived credentials that sit in their environment. The result is that the real access boundary lives in prompts and orchestration glue, which is exactly where it should not live.

This talk works through what identity for AI agents actually needs to look like: a distinct identity per agent, delegation that can only narrow as work is handed from one agent to the next, consent gates enforced by a signed grant the runtime checks rather than by instructions a prompt could override, credentials injected at call time so agents never hold secrets, tool servers pinned to the exact code you approved, revocation that takes effect mid task, and a tamper evident record of who did what and what was denied. We will ground each idea in a working open source implementation, Chancery, and connect it back to primitives the cloud native community already knows, including SPIFFE identity and software supply chain integrity.

Attendees will leave with a concrete checklist for evaluating least-privilege, delegation, and auditability in any multi-agent system, whether or not they use these particular tools.