AI & Agents

Agents at the Gate: Who Gets to Pull the Plug on a Bot?

Leave a comment

This was the week the agents ran into a locked door, a honeypot, and two different kill switches, and I kept coming back to the same question: when a bot can spend money, change an account, or wander off the rails, who is actually allowed to stop it?

United Airlines said it is banning AI agents from booking flights, CBS News reported, citing security risk and the chance that an agent checking you in could collide with rules that still expect a person to own the bags. To book, an agent needs credentials and a payment method. United’s point is that handing those to a third-party assistant can expose them, and can also produce purchases or MileagePlus changes nobody meant to make. I wrote that up when the ban landed.

The same week, Amazon’s threat intelligence team said AI agents had been hunting internet-exposed apps such as LiteLLM, Dify, and similar gateways for the API keys stored inside. The records, captured on decoy servers, describe stolen keys being used to run the models that searched for more keys. Amazon has moderate confidence the agents ran on ARTEX, and it cannot verify every outcome the agents claimed. The practical advice did not require a novel exploit write-up: change default administrator credentials, don’t leave an unfinished install on the public internet, apply the vendor updates, and rotate keys if you think a box was exposed.

Meanwhile the vendors who build and police these systems were talking about cutoffs. OpenAI has a system it calls Control Room, a kill switch for models in its research labs, The Register reported. It can halt a model’s output about 35 seconds after someone decides to pull the plug, a lag OpenAI blames on a caching layer, and it is designed to fail closed if Control Room itself goes down. CrowdStrike, earlier in this stretch of news, announced an AI Kill Switch in Falcon Guardian: a human-controlled brake that, today, works by cutting the models an agent depends on, with a roadmap to terminate the agent on the endpoint. Analysts were right to call it an emergency brake, not a rewind button. It does not undo what the agent already did.

My take

A bot that can book a ticket or hold a pile of API keys needs a human who can say no, and a way to make that no stick. United is drawing the line at the gate. Operators of exposed AI apps are finding out what happens when there was no line at all. Kill switches matter, but only if you know who is authorized to use them, what they actually stop today, and what they cannot take back. Least privilege still does more work than a red button you hope you never need.

God Bless, and Happy hacking!

References & Further Reading

Subscribe (RSS) Follow AI & Agents via RSS

Leave a note

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.