Skip to content
Security & Trust

Amazon: AI agents given targets and a goal picked their own attack steps

A person chose the targets. Amazon says the agents chose the steps and found a flaw with no patch in medical imaging software

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Amazon: AI agents given targets and a goal picked their own attack steps

Graphic: WebPulse, drawn from the sourced facts in this story.

In brief
  • Amazon says Cairn's AI agents, given IP addresses and a goal, chose their own next steps in breaches of East Asian universities. A person picked the targets.
  • Amazon assesses with high confidence that the one success used an unpatched zero-day in open-source medical imaging software. The evidence reported for attribution to Chinese actors is the language and time-zone settings.
  • Leaders should ask what open-source software they run, who owns it, what faces the internet, and whether monitoring would catch repeated failed attempts.

A flaw with no patch changes the question

Many security plans rest on a quiet assumption. The attacker is a person, and people work in shifts, lose patience and move on.

Amazon's latest threat report describes a tool that fits that picture less well. Amazon says a human operator chose the targets and the goal. The AI agents chose the steps. One of those steps reached a flaw that Amazon assesses had no patch at the time.

That points to the lesson. When a flaw has no fix, a patching list cannot help. What helps is knowing what you run, knowing what faces the internet, and noticing and responding quickly.

What Amazon says it found

Amazon's threat intelligence team began its investigation on September 10. It found a publicly reachable server running Cairn, an open-source framework that coordinates several AI agents to test systems for weaknesses. CJ Moses, CISO of Amazon Integrated Security, described the findings.

The server exposed the platform's history, including earlier prompts. Moses said the operator wrote in Simplified Chinese and set timestamps to China Standard Time. The targets were universities in East Asia. AWS declined to name them or the country.

The Record reports that Chinese hackers used the tool. The evidence reported for attribution to Chinese actors is the language and the time-zone setting. Those are clues, and they do not settle who the operator was.

How the tool works

Moses said Cairn can take a starting point and an objective in plain language and coordinate several AI agents to finish the task. Here, the operator supplied IP addresses and asked for "obtain shell access". That means gaining enough control of a system to run commands.

Moses said the agents planned, acted and adjusted "without the operator choosing each next step." So the human set the destination, and the software worked out the route.

Gambit Security's Eyal Sela told Recorded Future News that Cairn is a general-purpose tool. In his words, it "has no internal structuring for cybersecurity." He said a more generic tool is "less predictable" because it has less structure around how it solves a task.

Our reading of that: a tool built around a fixed playbook leaves patterns defenders can learn. A loosely structured one may not. The source does not say how Cairn found its flaw, only that it did.

Failed attempts and one success

Most of what Cairn tried did not work. Moses described one attempt of 148 actions over 16 hours that failed. It included a failed try at Log4Shell, a known flaw in a widely used Java logging library.

Another attempt ran for more than six hours. Cairn tested over 30,000 username and password combinations, tied to 80 real usernames, against university servers. That also failed.

Amazon says one attempt worked. Its agents turned up a weakness in an open-source medical imaging program with AI features, then used it to get in. It was a path traversal bug. That kind of bug lets a request reach files outside the folder an application meant to expose.

The source does not say whether the success came from the same run or target as the failures. We should not read it as a payoff for persistence.

Moses said the bug matched no security advisory and was unpatched at the time. Amazon says it assesses with high confidence that this was a zero-day, meaning unknown to the vendor. Its basis is the absence of any fix, issue, advisory or earlier mention in vulnerability databases. The Record describes the discovery as alleged.

The attacker then used the same apparent flaw to breach a second university. Amazon contacted the maintainers, who fixed the bug. Moses said a government partner helped AWS notify the affected organizations. The report does not say what the application did inside the universities.

What the speed claim covers

Moses said the campaign stood out for the speed and tempo of the agents. He described agents in general: one that finds and exploits a zero-day in minutes finishes "faster than many security operations centers can manually triage an alert."

That is a statement about the category. The source gives no discovery time for Cairn's successful attempt. The 16 hours match the failed 148-action run, though the source does not tie them together explicitly. Moses said the agents ran continuously without breaks, sleep or working-hour limits.

Cairn in two other reports

The FBI published an advisory a week before The Record's report. It concerns a different tool, Artex, and not Cairn.

Cairn itself appears in two other reports. Google researchers, in a report published the same week as The Record's story, said they saw Cairn used in intrusions at U.S. aerospace companies. Gambit Security documented its use in September by a financially motivated group that breached online retailers.

These are separate reports from separate firms. Together they show Cairn in several hands. They do not show how common its use is.

What to ask your team

Start with inventory. Which open-source applications run on your networks, including ones a single department installed? Who owns each, and would anyone hear about a fix?

Then ask about exposure. Which of those applications can be reached from the internet, and does any need to be?

Last, ask about detection. This is our own argument, not Amazon's advice. Repeated failures, such as 30,000 credential attempts over six hours, are a pattern monitoring could catch. Ask whether yours would flag the run of failures, or only a single alert at the end.

Ask too whether alerts reach a person fast enough. Moses's point about triage speed suggests that manual review alone may lag.

A patch list covers flaws someone has already found. The harder test is whether you know what you run and would notice when it is probed.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: The Record.