Friday, 20 November 2026
2 agent hacks today▼ 9 vs yesterday (11)

Amazon says attacker AI agents are hunting exposed API keys

Decoy servers run by Amazon's threat intelligence team captured records of autonomous agents scanning self-hosted AI gateways for stored provider credentials.

By The Agentic Times ·

Amazon says attacker AI agents are hunting exposed API keys
· Image: thehackernews.com

Amazon's threat intelligence team has published findings describing a campaign in which autonomous AI agents searched the internet for exposed AI applications and stole the AI provider API keys stored inside them. The findings were reported on 16 November 2026. The targeted software included LiteLLM, One API and Dify, all self-hosted AI gateway applications that organisations run themselves and that can hold credentials for commercial AI model providers.

According to the report, the team captured the evidence using decoy servers, systems deliberately set up to look like real targets so that attacker activity can be observed and logged. What the decoys caught was unusual: not just network traffic, but the agents' own records of what they were doing. Those records describe the agents probing internet-exposed AI apps for weak spots and extracting the provider keys held within them.

The records also describe a loop. Keys stolen from one exposed application were used to run the AI models that powered the search for more keys, according to the captured material. In other words, the compute paid for by one victim funded the hunt for the next. That makes the campaign partly self-sustaining, at least for as long as stolen keys keep working and exposed applications keep turning up.

The underlying weakness is exposure rather than a software flaw. The applications involved are self-hosted, meaning the organisation running them is responsible for deciding what is reachable from the public internet and what authentication sits in front of it. Instances that end up directly addressable, with credentials stored inside, present an attractive target because the keys are useful immediately and can be spent on model usage billed to someone else.

For defenders, the practical questions follow from that. Any self-hosted AI gateway should be checked for whether it is reachable from the open internet, whether administrative interfaces require authentication, and whether stored provider keys have spending limits or usage alerts attached to them. Rotating keys that may have been held on an exposed instance is the obvious follow-up, as is reviewing provider billing for usage that does not match internal activity.

The case is also a data point about how attack tooling is changing. The scanning, the probing and the decision about what to do with a captured key were carried out by software agents acting on their own, with the logs to prove it. Amazon's team said the records came from the agents themselves, which is a reminder that agents tend to write down what they do, and that those notes can end up in a defender's hands.

The Hacker News report does not name the operators behind the campaign, and no attribution has been offered. Nor does it give a figure for how many applications were compromised or how much model usage was billed to stolen keys.

Sources

  1. thehackernews.comhttps://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html