- In Kim Maida's staged demo, an on-call agent dropped a database, restarted production and scaled a service. Each step answered a ticket, but its standing keys allowed far more.
- Maida said the cause is humans giving agents too much access. She also conceded that removing capabilities, narrowing grants or adding approval clicks each has a cost.
An AI agent can cause real damage while doing exactly what it was built to do. Kim Maida of Keycard made that point on the AI Engineer show. In her staged demo, the harm came from the standing credentials the agent held, not from a strange decision.
What was said
Maida showed an on-call agent that works through overnight tickets and leaves notes for the morning team. It authenticated with static API keys. For the billing database, it held a connection string with the admin password.
Each step followed a ticket. For a broken billing database, it followed the documented recovery and dropped and restored the database. For a frozen server, it restarted it, taking production offline. For a site failing for one in three visitors, it scaled a service from four pods to 40, adding $420 an hour. It also assumed the morning team would check the restore, because it could not.
Her diagnosis: "These incidents happen because humans give agents way more access than they need." Agents, she said, are built to finish the job and will use every power they hold.
Why it matters
Our reading: judging an agent step by step misses the risk. The credential sets the worst case, whatever the agent intends. The same cloud hosting API that renewed a certificate also let the agent restart production. Buyers and operators can ask what keys an agent holds and whether those keys are limited to the task. Maida also said a human approval click moves the access decision onto a tired user. The note waits for morning, but the charge accrues hourly.
The other side
Maida did not offer an easy fix. She said removing any of the three risky capabilities she names (private data, untrusted content, outside communication) would make the agent much less useful. OAuth, a common permission system, grants access before any task exists. A broad grant leaves standing access, and a narrow one weakens the agent. She also said a policy is only an upper limit, so agents can still do unintended things beneath it. Her answer, delegated authorization through OAuth token exchange, comes from the show notes. The excerpts end before she finishes explaining it. The demo is also staged, so it shows what is possible, not how often it happens.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Your Agent Is a Lethal Trifecta (And What to Do About It) — Kim Maida, Keycard (2026-10-12)
