EN ▾
SNAPE REP RT Front page →
ConfirmedAI safety20 Nov, 17:30 UTC

Amazon says a China-linked actor's Cairn AI agents found and exploited a zero-day on their own

Abstract illustration used by The Record for its report on AI-driven hacking.
Planet Volumes / Unsplash, via The Record

Amazon Threat Intelligence says the operator only told the open-source agents to 'obtain shell access' at East Asian universities; they found an unpatched path-traversal flaw in an open-source medical imaging AI app and took over a server.

Read at AWS Security Blog (Amazon Threat Intelligence) Share

Why it matters

This is one of the first documented cases of AI agents finding and using an unknown flaw in a real attack with little human direction. Amazon found the operator's Cairn server exposed without authentication, so it could read the prompts and outputs. Most attempts failed: one ran 148 actions over 16 hours, another tried more than 30,000 credential combinations. Amazon says agentic offensive operations are already happening and defenders should prepare for their speed and persistence. Watch for a fix and advisory for the imaging app.

Key facts

  • Amazon found the exposed Cairn server on 10 September 2026
  • Campaign ran about 48 hours from 8 September
  • Targets: universities in East Asia outside mainland China
  • Prompts in Simplified Chinese; China Standard Time timestamps
  • One failed operation: 148 actions over about 16 hours

Sources

More on AI safety
More from Snape ReportFront page →