Skip to content

Amazon: autonomous AI tool Cairn found a zero-day in a live attack

Amazon says a China-nexus actor used Cairn, an open-source autonomous AI pentest tool, to find and exploit a path-traversal zero-day during attacks on East Asian universities.

Published 4 min read

Amazon Threat Intelligence published a case study worth an operator's attention: an attacker pointed an open-source autonomous AI penetration-testing framework at live targets, and it found and exploited a genuine zero-day. The tool is called Cairn, and Amazon's writeup is unusually concrete because investigators recovered the operator's prompt history from an exposed server. Strip away the framing and the useful signal is mixed — mostly noise, one real win — which is exactly why it's worth reading carefully rather than at headline speed.

What Cairn is

Cairn is described as an open-source framework that coordinates multiple AI agents from a plain-language objective: give it a starting point and a goal, and it plans and executes steps on its own. It is not bespoke nation-state tooling — Google has separately observed Cairn in intrusions against US aerospace firms, and Gambit Security reported in September that a financially motivated group used it to breach online retailers. In other words, it's a commodity capability, available to whoever wants it.

What Amazon observed

Amazon's investigation began on September 10, when its team found a publicly accessible server running a Cairn instance — and could read the operator's prompt log. According to Amazon (CJ Moses, CISO of Amazon Integrated Security), the operator handed Cairn a set of IP addresses and asked it to "obtain shell access." Over roughly 16 hours the tool took about 148 actions, including a Log4Shell exploit attempt and more than 30,000 credential combinations against 80 usernames across six-plus hours.

Almost all of it failed. The one success was a path-traversal zero-day in an open-source medical-imaging AI application — the linked GitHub advisory points to Project-MONAI/MONAILabel — which the tool then used against a first university and a second. Amazon says it found no matching advisory, issue or prior database entry, and assesses "with high confidence" that it was a zero-day. The maintainers were contacted and fixed the bug; no CVE is cited in the reporting.

The targets were several universities in East Asia, which Amazon declined to name. AWS refers to "Chinese nexus actors"; that attribution is hedged — the cited evidence is prompts written in Simplified Chinese and timestamps set to China Standard Time, both of which are suggestive, not dispositive.

The limits worth keeping

Read the numbers honestly. The autonomy is real, but the hit rate was low: thousands of failed credential guesses, a Log4Shell attempt against presumably-patched infrastructure, and exactly one novel exploit — against a niche medical-imaging app, not a hardened enterprise edge. The notable part isn't that AI "discovered a zero-day" like a magic oracle; it's that an off-the-shelf agent brute-forced its way through a target list unattended and surfaced one path-traversal bug among the noise. That lowers the skill floor for opportunistic attackers more than it raises the ceiling for sophisticated ones.

What it means for defenders

  1. Niche internet-exposed apps are now in scope. A medical-imaging AI tool with a path-traversal bug is precisely the long-tail asset these agents will enumerate and probe. Inventory the obscure web apps on your perimeter.
  2. Volume detection still works. 30,000 credential attempts and a Log4Shell probe are loud. Rate-limiting, lockouts and WAF signatures catch the bulk of what an unattended agent throws.
  3. Patch the long tail. The agent's only win was against unpatched/unknown code. Dependency and asset hygiene remain the control that matters.
  4. Expect commodity autonomy. Cairn is open source and already used by multiple unrelated actors. Plan for "cheap, tireless, mediocre attacker" as a baseline, not an exception.

This is the AI-security beat at its most grounded: not a demo, but a logged, real-world run — and the honest reading is that it automated the boring 90% of an intrusion and got lucky once. That's still a meaningful shift, and it's a more useful lesson than either the hype or the dismissal.

Related stories