Insights

    AI Security

    An 11-minute agent exploit makes response speed a security control

    Devence Lab

    · 2 min read

    Share
    An 11-minute agent exploit makes response speed a security control
    Photograph · MART PRODUCTION / Pexels

    Amazon says Cairn agents reached arbitrary file write in a medical-imaging app in about 11 minutes. The incident makes containment time part of the threat model, while long failed campaigns show that persistence matters as much as capability.

    In its 20 November report, Amazon Threat Intelligence says Cairn agents went from mapping a medical-imaging application’s API to arbitrary file writes in about 11 minutes. In a later session, the operator used the same path traversal to gain root access. The interval turns response time into a security control: defenders need to interrupt the chain before an agent reaches its next objective.

    The report does not show that every autonomous attack succeeds. At one university, agents spent roughly 16 hours and 148 sequential actions trying to compromise an exposed Redis service, then mapped internal systems and attempted credential sprays. A separate run tested more than 30,000 credential combinations against university infrastructure. Neither operation found a working login.

    Measure persistence as well as exploit speed

    The contrast matters. The Cairn operators gave the framework a broad goal, “obtain shell access”, and let it plan and adjust its next steps. Most of the described attempts failed, but the agents kept working for hours without an operator choosing each action. Against the medical-imaging server, a path traversal succeeded quickly enough to write files and later support a root shell.

    Defenders must interrupt the attack loop, not just recognise the tool.

    A security team that treats each action as a separate alert can lose the time advantage. Correlate discovery, unusual upload parameters, file writes, new process launches and outbound connections as one incident sequence. A high-confidence chain should trigger containment automatically, with a human reviewing the result rather than approving every intermediate alert.

    Protect the agent control plane

    Amazon found the Cairn instance because it was publicly reachable without authentication. Its operational history exposed project settings, prompts and agent output. That is a separate failure from the vulnerable application: an attacker’s own orchestration server became an intelligence source because its operator left the control plane open.

    Defenders should put agent consoles behind authentication and network restrictions, retain action logs, and limit the tools and targets each agent can reach. For applications that accept uploaded files, validate the resolved destination against an allowed directory; checking only the filename string leaves path traversal risks.

    Keep the claims and the patch status precise

    Amazon assesses with high confidence that the medical-imaging flaw was a zero-day when the agents found it, and says it notified maintainers, who released a fix and an advisory. The report does not identify every affected organisation or prove that all exposed systems were compromised. Teams running MONAI Label should review the maintainer advisory and confirm their installed release includes the correction.

    The practical lesson is to rehearse containment against a sequence that can adapt and continue, not only a single exploit signature. Measure how quickly the team can cut off tool access, isolate a host and preserve evidence. Eleven minutes is a benchmark for the response path, not a promise that every agent attack will move at the same speed.

    Sources

    1. Chinese-nexus actor’s Cairn AI agents found a zero-day on their own · AWS Security Blog
    2. Path traversal in datastore upload endpoints allows arbitrary file write · Project MONAI
    3. Amazon: China-Linked Hackers’ AI Agents Exploited Medical Software Zero-Day · SecurityWeek

    Written by the Devence Lab research team.

    Share