Uploaded October 2026 | Updated November 2026, 4 days ago
An on call agent drops a billing database, restarts a production server and scales a service from four pods to 40, adding $420 an hour to the bill. Kim Maida's staged triage demo shows what happens when an agent receives standing API keys and an administrator's database credentials. Private data, untrusted tickets and access to external services combine into the lethal trifecta. The agent can do useful work, but its credentials give it far more authority than any one ticket requires. A human approval button alone cannot tell it which actions the user is actually authorized to approve.
Maida walks through delegated authorization using OAuth token exchange: carry the user's identity through the agent and MCP server, evaluate policy at a security token service and issue short lived credentials scoped to the requested access. She also explains Keycard's approach to intent based authorization, which she identifies as a proposal rather than a finalized standard. The second demo uses task based policy instead. Customer data access and database deletion are denied, certificate renewal is allowed, and a server restart stays blocked even after approval because the user's role cannot authorize it. Scaling proceeds when the user has the required authority. The demos make identity, delegation and policy decisions visible throughout the request chain, while keeping authorization distinct from whether an agent's answer is correct.
Speaker info:
- linkedin.com/in/kimmaida
- github.com/kmaida
- https://maida.kim/
Timestamps:
00:00 - An overprivileged on call agent
2:32 - Database deletion, restart and scaling
4:16 - The lethal trifecta
6:21 - Access control and standing credentials
8:43 - Missing identities and token exchange
10:25 - Following the agent and MCP execution path
12:18 - Subject tokens and the security token service
14:11 - Intent based authorization as a proposal
16:05 - Ephemeral credentials and delegation chains
17:37 - Repeating the demo with task based policy
20:01 - Data access, renewal and deletion decisions
22:28 - Approval constrained by the user's authority
---
Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page:
https://ai.engineer/talks/ws9XcGGPBDk-youtube-aa04df8634861a97a162
Join our upcoming AI Engineer conferences (see https://ai.engineer landing page):
- AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers
- AIE CODE in SF, November 10–12 — the top AI Coding conference returns
- AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney
---
An on call agent drops a billing database, restarts a production server and scales a service from four pods to 40, adding $420 an hour to the bill. Kim Maida's staged triage demo shows what happens when an agent receives standing API keys and an administrator's database credentials. Private data, untrusted tickets and access to external services combine into the lethal trifecta. The agent can do useful work, but its credentials give it far more authority than any one ticket requires. A human approval button alone cannot tell it which actions the user is actually authorized to approve.
Maida walks through delegated authorization using OAuth token exchange: carry the user's identity through the agent and MCP server, evaluate policy at a security token service and issue short lived credentials scoped to the requested access. She also explains Keycard's approach to intent based authorization, which she identifies as a proposal rather than a finalized standard. The second demo uses task based policy instead. Customer data access and database deletion are denied, certificate renewal is allowed, and a server restart stays blocked even after approval because the user's role cannot authorize it. Scaling proceeds when the user has the required authority. The demos make identity, delegation and policy decisions visible throughout the request chain, while keeping authorization distinct from whether an agent's answer is correct.
Speaker info:
- linkedin.com/in/kimmaida
- github.com/kmaida
- https://maida.kim/
Timestamps:
00:00 - An overprivileged on call agent
2:32 - Database deletion, restart and scaling
4:16 - The lethal trifecta
6:21 - Access control and standing credentials
8:43 - Missing identities and token exchange
10:25 - Following the agent and MCP execution path
12:18 - Subject tokens and the security token service
14:11 - Intent based authorization as a proposal
16:05 - Ephemeral credentials and delegation chains
17:37 - Repeating the demo with task based policy
20:01 - Data access, renewal and deletion decisions
22:28 - Approval constrained by the user's authority
---
Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page:
https://ai.engineer/talks/ws9XcGGPBDk-youtube-aa04df8634861a97a162
Join our upcoming AI Engineer conferences (see https://ai.engineer landing page):
- AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers
- AIE CODE in SF, November 10–12 — the top AI Coding conference returns
- AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney
---










