Hackers used an autonomous AI tool called Cairn to find a Cairn zero-day in medical imaging software and take over a server, a milestone for AI-driven offensive security.
According to BleepingComputer's report, Amazon Threat Intelligence uncovered the campaign after the attackers left their own Cairn instance publicly accessible without authentication. That accident of operational security is the only reason the outside world can see what autonomous AI agents actually do when a human operator simply tells them to "obtain shell access." For healthcare organizations, universities, and defense contractors that run AI-adjacent open-source software, this is the first well-documented look at the technique end to end, and it deserves close reading rather than a shrug.
Key Takeaways
- A Chinese-nexus threat actor, attributed with moderate confidence, used Cairn, a legitimate AI pentesting platform, between September 8 and 10, 2026, targeting mainly East Asian educational institutions outside China.
- The AI agents were given a simple goal, "obtain shell access," and independently discovered and exploited a path traversal zero-day in MONAI Label, an open-source medical imaging application.
- From mapping the application's API to exploiting the vulnerability took 11 minutes, and the agents used the flaw to write a Python reverse shell that gave them root access to the server.
- Autonomy is not invincibility: a 16-hour, 148-action attempt against a Redis server failed, and agents tried over 30,000 credential combinations over six hours against a university GPU server without success.
- MONAI Label maintainers released a security update on November 18 after Amazon reported the flaw, and the underlying path traversal was fixed by validating that uploaded files stay inside the datastore directory.
What Happened: The Facts in Detail
How the campaign came to light
Amazon Threat Intelligence discovered the operation because the attackers made a classic mistake: they exposed their Cairn instance to the internet without authentication. Cairn is a legitimate pentesting tool created for research and defense strengthening. It can coordinate multiple AI agents in cyber-operations. Because the operators left their instance open, researchers could retrieve the prompts, observe the agent activity, and reconstruct the campaign in unusual detail.
Who was behind it and when
The researchers attributed the campaign to a Chinese-nexus threat actor with moderate confidence. The attribution rests on three signals: the language of the retrieved prompts, the activity timestamps, and the choice of targets. The attacks ran between September 8 and 10, 2026, and targeted mainly East Asian educational institutions outside China. Amazon also shared two IP addresses associated with the campaign, observed in that same window.
Moderate confidence is an honest label, and readers should treat it that way. The evidence is behavioral and linguistic, not a smoking gun such as seized infrastructure. That said, the operational facts, the tooling, the timing, and the exploitation chain, stand on their own regardless of who pressed the button.
The target: MONAI Label
The AI agents found a path traversal zero-day in MONAI Label, an open-source medical imaging application that uses AI to help researchers label medical images. The project's own security advisory explains the mechanics. The PUT /datastore/image endpoint accepts a caller-controlled image query parameter that is used as the image ID without validation. The application then appends a file extension to that ID, joins it onto the datastore image directory, and writes the uploaded file to the resulting path. Because the resolved path is never checked to be inside the datastore, a value such as ../../../tmp/x writes uploaded content to any location the MONAI Label process can write to. The PUT /datastore/label endpoint builds its destination the same way.
Two configuration facts made this worse. Authentication is disabled by default, so in the default configuration any network client that can reach the server can exploit the flaw. Even with authentication enabled, the ANNOTATOR role is sufficient to trigger it. An arbitrary file write of this kind can lead to remote code execution, for example by overwriting application code, adding entries to .ssh/authorized_keys, or writing cron entries, or to denial of service by overwriting configuration files.
The exploitation chain, minute by minute
Within 11 minutes, the agents went from mapping the application's API to exploiting the vulnerability. In a subsequent session, they used the flaw to write a Python reverse shell into sitecustomize.py, a file that automatically loads when Python starts. Then they triggered an API endpoint to establish root access to the server. Read that sequence again if it has not sunk in: the agents found an unpatched flaw in software their operator did not point them at, weaponized it, achieved persistence through a Python startup hook, and escalated to root, all without step-by-step human direction.
After discovering the zero-day, the operator shifted from observer to director. They provided the agents with a proxy into a second university network running the same vulnerable software and directed them to target other machines. The human contribution was targeting and access; the technical discovery and exploitation work belonged to the agents.
The failures matter as much as the success
The campaign was not a clean sweep, and Amazon's reporting is candid about that. In one case, a 16-hour and 148-action effort to take control of a Redis server failed. In a separate case targeting a university GPU server, the agents tried over 30,000 credential combinations over six hours and failed to brute-force their way in. These failures are analytically important for two reasons. First, they show that autonomous agents still stall against properly hardened services. Second, they show the economics of machine-speed attack: 30,000 credential attempts in six hours is a rate of effort no human operator sustains, and the fact that it failed says something useful about what defensive controls actually stop.
Disclosure and fix
Amazon reported the flaw to the MONAI Label maintainers, and the maintainers released a security update to address it on November 18. Per the advisory, the fix resolves the destination path, including the label tag directory, and rejects any path outside the datastore image and labels directories. The /datastore/image endpoint also reduces the image ID to its basename and rejects empty or dot-prefixed IDs with an HTTP 400 response. The advisory's recommended mitigations are concrete: enable authentication, do not grant the ANNOTATOR role to untrusted users, do not expose the MONAI Label server to untrusted networks, and run the server as an unprivileged user with write access limited to the datastore directory.
The Cairn Zero-day Chain: How It Works and Why It Matters
Strip away the novelty and the Cairn zero-day attack is a chain of three well-understood weaknesses, assembled at machine speed. The first weakness is an unvalidated input that becomes a filesystem path. The second is a default-off security control, authentication, that most research deployments never switch on. The third is excessive privilege: a process running with enough rights that an arbitrary file write becomes root code execution. None of these is exotic. What is new is that an autonomous agent found the first, chained it through the second and third, and did so in 11 minutes without a human writing an exploit.
The persistence technique deserves particular attention from IT leaders. The agents did not drop a binary in a temp folder or install a service, the patterns most endpoint tools are tuned to flag. They wrote a Python reverse shell into sitecustomize.py, which executes automatically whenever Python starts. Any environment that runs Python, which describes most AI research, medical imaging, and data science infrastructure, will re-trigger that payload routinely. Detection here is less about signature matching and more about integrity monitoring of startup paths and egress control on outbound connections.
The operator's role in this campaign also reframes what "AI attack" means. The human did not write the exploit, did not find the vulnerability, and apparently did not know in advance which flaw the agents would use. The human set an objective, supplied network access, and redirected the agents toward a second victim. That division of labor lowers the skill floor for offensive operations dramatically. It also means the defenders' problem changes shape: you are no longer looking for a specific exploit or a specific actor's tooling, you are looking for any process that maps your APIs, probes your endpoints, and writes files outside expected directories, at any hour, at machine pace.
Amazon's warning follows directly from that observation. AI agents can find and exploit vulnerabilities faster than security teams can react, and the company urged organizations to rely more on automated defenses rather than manual investigations. For readers thinking about their own AI-adjacent infrastructure, the practical translation is that the MONAI Label advisory's mitigations, authentication on by default, no exposure to untrusted networks, least-privilege service accounts, and constrained write paths, are exactly the controls that would have broken this chain at multiple points. Our overview of managed cybersecurity services covers how layered detection and response maps to attack patterns of this kind, and our analysis of securing autonomous AI agents in the enterprise addresses the defensive mirror image of this story: the same agent architectures that attackers abused can be run under zero-trust controls on the defender's side.
What the Evidence Does and Does Not Show
Strong reporting still has limits, and regulated readers should calibrate on them before changing budgets or board narratives.
First, the attribution is moderate confidence, based on prompt language, timestamps, and target selection. Amazon has not named a specific group, and nothing in the public reporting establishes state sponsorship or a broader program. Treat "Chinese-nexus threat actor" as a characterization of the evidence, not a settled conclusion.
Second, the campaign window is narrow: September 8 through 10, 2026, against mainly East Asian educational institutions. The public record does not establish how many organizations were targeted, how many were compromised, or whether data was exfiltrated from the root-accessed MONAI Label server. The reporting describes server takeover, not theft of patient data or research data, and no victim counts appear in the sources.
Third, the discovery of the campaign depended on the attackers' own misconfiguration, an unauthenticated, publicly exposed Cairn instance. Campaigns run by operators who secure their tooling will not be visible this way. It is reasonable to assume, as Google has in warning that fully-automated AI attacks are coming, that this technique is being refined out of public view. It is not reasonable to assume this is the only instance.
Fourth, the failure cases are real but bounded. The agents failed against a Redis server after 16 hours and 148 actions, and failed to brute-force a university GPU server after over 30,000 credential combinations in six hours. That is evidence that hardening works, not evidence that hardening is optional. The same campaign succeeded against a default-configured MONAI Label deployment in 11 minutes. The difference between the outcomes is configuration and exposure, not luck.
Fifth, some technical questions remain open. The sources do not state which AI models powered Cairn's agents, how the operators prompted or constrained them beyond the "obtain shell access" instruction, or whether the agents' zero-day discovery was a one-off or repeatable across other codebases. They also do not describe any detection that caught the attack in progress; the campaign came to light through the exposed instance, not through a victim's security operations. That gap, between exploitation and detection, is arguably the most important finding of all, and it is the thread we pick up in Part 2.
For organizations that want the underlying detail, the primary materials are the BleepingComputer report and the MONAI Label advisory. Readers planning defensive work can start with CISA's Cross-Sector Cybersecurity Performance Goals and NIST SP 800-61 Rev. 3 on incident response, both of which frame the baseline practices that this attack chain would have tested.
Compliance and Risk Implications for Regulated Organizations
The Cairn campaign is not just a novel attack story. It is a stress test of the assumptions that most compliance programs are built on, and several of those assumptions did not hold up well.
The first broken assumption is that zero-day exploitation requires a nation-state budget and weeks of operator effort. Here, a legitimate pentesting platform was pointed at a target with a one-line instruction to obtain shell access, and the agents moved from mapping the MONAI Label API to exploiting a path traversal flaw in 11 minutes. Amazon Threat Intelligence also documented the failures: a 16-hour, 148-action attempt against a Redis server that did not succeed, and more than 30,000 credential combinations tried over six hours against a university GPU server without success. The lesson is not that AI agents always win. It is that the cost of attempting exploitation has collapsed to nearly zero, so the volume of attempts your defenses face is going up even when the success rate stays low.
The second broken assumption is that a vulnerability in an open-source research or clinical tool is a low-priority patch. MONAI Label is a medical imaging application that uses AI to help researchers label medical images, and its default configuration ships with authentication disabled, meaning any network client that can reach the server can exploit the flaw, as the maintainers' advisory explains. The maintainers released a security update on November 18 after Amazon reported the flaw. If your organization runs AI-adjacent research or clinical software, it belongs in your asset inventory and your patch cycle, not in an unmanaged corner of a lab network.
For defense contractors, the connection to NIST SP 800-171 is direct. The incident response family of that framework requires an operational incident handling capability, with incidents tracked, documented, and reported. An attack that moves from initial access to root on a server in minutes is exactly the scenario where a manual, ticket-driven investigation process fails. DFARS 252.204-7012 separately requires reporting a cyber incident affecting covered defense information to DoD within 72 hours of discovery, with system images and monitoring data preserved for at least 90 days afterward. Detection speed is now a compliance requirement, not just a security preference.
For healthcare organizations, the HIPAA Security Rule requires a risk analysis under 45 CFR 164.308(a)(1)(ii)(A), and a risk analysis that has never considered AI-driven offensive tooling is incomplete. IBM's 2026 figure puts the average healthcare breach at $6.64M, and the sector's breach volume, including 725 large healthcare breaches in 2024, shows how much exposure accumulates when research and clinical systems sit outside the main security perimeter.
For anyone handling payment card data, PCI DSS v4.0.1's 12 control areas assume attackers operate at human speed. Autonomous agents compress the timeline between reconnaissance and exploitation, which raises the bar on continuous monitoring and automated response. CISA's Cross-Sector Cybersecurity Performance Goals 2.0, aligned to the NIST Cybersecurity Framework 2.0 functions including GOVERN, offer a practical baseline for organizations that need to prioritize limited investment, and NIST SP 800-61 Rev. 3 frames incident response as something to prepare for before the incident, not after.
What This Means for Regulated Industries
Defense Contractors
The Cairn operators were attributed with moderate confidence to a Chinese-nexus threat actor, based on the language of the retrieved prompts, activity timestamps, and target selection. Defense contractors are squarely in the set of organizations that Chinese-nexus actors target, and the tooling demonstrated here works against any internet-reachable server running vulnerable software. Three actions matter most. First, confirm your NIST SP 800-171 self-assessment score is current in SPRS, since DFARS 252.204-7019 requires it. Second, verify your incident reporting path works in practice: the DFARS 252.204-7012 72-hour reporting window requires a DoD-approved medium assurance certificate obtained in advance, not requested during an incident. Third, treat any AI or research computing environment holding CUI as in scope for your assessment boundary. Enclaving CUI into a narrowly scoped boundary reduces audit surface and is one of the biggest drivers of both cost and risk reduction.
Healthcare
The exploited application, MONAI Label, is medical imaging software. That should get the attention of every hospital research division and imaging informatics team. The advisory's own mitigations are a checklist worth adopting immediately: enable authentication, do not grant the annotator role to untrusted users, do not expose the server to untrusted networks, and run it as an unprivileged user with write access limited to the datastore directory. Beyond that specific flaw, healthcare organizations should assume that AI-assisted research tools are attractive targets precisely because they sit on networks with PHI nearby and are often excluded from enterprise patching. The HIPAA Breach Notification Rule at 45 CFR 164.404, 164.406, and 164.408 sets your notification obligations, but only detection and logging make those obligations achievable on an AI-speed attack timeline.
Legal
Law firms hold client confidences that carry contractual and, in many matters, regulatory weight, and they increasingly run AI tooling for discovery and document review. The Cairn campaign shows that the AI tools themselves, and the servers they run on, are attack surface. A GPU server used for AI workloads is a high-value target: the attackers in this campaign specifically went after university GPU servers. Firms should inventory every AI system and supporting compute resource, confirm who can reach them over the network, and apply the same access controls and logging to them as to any system holding client data. The 30,000-credential brute-force attempt against a GPU server is a reminder that these systems face credential attacks at machine speed, which makes MFA and account lockout non-negotiable.
Financial Services
Financial institutions already operate under continuous monitoring expectations, so the gap here is narrower but real: the speed of the attack chain. An 11-minute path from API mapping to exploitation means quarterly penetration tests and monthly patch cycles leave long windows. Gambit Security documented Cairn abuse in a large-scale credit card stealing operation in September, and Google reported it deployed against a U.S. aerospace organization, which tells you the tooling is already spreading across criminal and state-aligned use cases. Institutions should prioritize automated detection and response over manual investigation workflows, consistent with Amazon's own warning that AI agents can find and exploit vulnerabilities faster than security teams can react.
An Action Plan for This Month
- Inventory your AI and research computing systems. List every server running AI-adjacent software, including medical imaging, labeling, and model training tools, along with who can reach them over the network. The Cairn campaign succeeded against a system nobody was watching.
- Check for the MONAI Label fix specifically. If your organization runs MONAI Label, apply the security update the maintainers released on November 18, and review the maintainers' advisory for the full mitigation list.
- Close default-configuration exposure. The MONAI Label flaw is exploitable by default because authentication ships disabled. Audit your other open-source and internal tools for the same pattern, and never expose them to untrusted networks.
- Run vulnerable services unprivileged. Apply the advisory's guidance to limit write access to the minimum needed. The Cairn agents achieved root by writing a Python reverse shell to a file that loads automatically; a constrained service account breaks that chain.
- Test your detection timeline. Ask how long it would take your team to notice reconnaissance and exploitation activity on a lab or research server. If the honest answer is days, that is the gap autonomous attackers exploit.
- Verify your incident reporting readiness. Defense contractors should confirm the DFARS 252.204-7012 reporting path, including the medium assurance certificate, is in place before an incident. Healthcare organizations should rehearse breach notification triage under 45 CFR Part 164.
- Deploy MFA and rate limiting on all remote and administrative access. The 30,000-credential attempt against a university GPU server failed, but only because the agents ran out of patience, not because the target was inherently resistant.
- Preserve evidence capability. DFARS 252.204-7012 requires system images and monitoring data preserved for at least 90 days after a reported incident. Confirm your logging retention actually supports that.
- Reassess your risk analysis for AI-speed threats. Update your HIPAA risk analysis, NIST SP 800-171 gap assessment, or PCI DSS control review to explicitly consider autonomous attack tooling, using CISA's Cross-Sector Cybersecurity Performance Goals 2.0 as a prioritization baseline.
- Move from manual to automated defense. Amazon's guidance is explicit: organizations should rely more on automated defenses rather than manual investigations. If your response model depends on a human reading an alert first, rework it.
How Petronella Technology Group, Inc. Helps
The countermeasure to machine-speed attack is machine-speed defense, and that is where our practice is built. Our managed detection and response service combines 24/7 threat monitoring, AI-powered analytics, and human-led investigation, replacing the need for an in-house security operations center. Our hybrid SOC runs ten-plus production AI agents on our own private AI cluster, serving defense industrial base and healthcare clients that cannot send CUI or PHI to a public-cloud SOC. The guardrails matter as much as the capability: the AI never closes a ticket on its own, never touches production systems without human authorization, and every action is logged for CMMC and HIPAA audit. Onboarding is operational within 30 days of contract signing, and reporting is designed to give auditors the evidence they need for HIPAA, CMMC, PCI DSS, SOC 2, and NIST 800-171 verification.
For organizations that need to adopt AI themselves without creating the next unmonitored GPU server, we design, build, and operate private AI clusters on a segmented VLAN or full air gap, with role-based access control, encryption, audit logging mapped to framework controls, and validation before production. Our seven-stage method starts by defining the data boundary, whether that is CUI, PHI, or financial records, and ends with controls validated against CMMC, HIPAA, or DFARS 252.204-7012 requirements.
On the compliance side, our compliance practice maps your current state to the frameworks that apply to you, ranks gaps by priority, and produces a remediation roadmap written for the executive reader. Every practitioner on the team holds the CMMC-RP designation, and we are a Cyber AB Registered Provider Organization, RPO #1449. Related reading from our blog includes Zero-Trust AI: How to Secure Autonomous Agents in the Modern Enterprise and AI-Driven SOC: Autonomous Threat Response, both of which go deeper on defending against and deploying agentic AI safely.
Sources
- Hackers use autonomous AI tool Cairn to find zero-day, take over server
- Path traversal in datastore upload endpoints allows arbitrary file write · Advisory · Project-MONAI/
- Cross-Sector Cybersecurity Performance Goals
- SP 800-61 Rev. 3, Incident Response Recommendations and Considerations for Cybersecurity R
Frequently Asked Questions
What is the Cairn zero-day vulnerability?
It is a path traversal flaw in MONAI Label, an open-source medical imaging application that uses AI to help researchers label medical images. The PUT datastore upload endpoints accepted a caller-controlled image parameter without validating the resulting file path, allowing an attacker to write files to arbitrary locations, which can lead to remote code execution. The maintainers released a security update on November 18 after Amazon reported the flaw.
How did hackers use AI to exploit the zero-day?
According to Amazon Threat Intelligence, a Chinese-nexus threat actor instructed Cairn's autonomous AI agents to obtain shell access, and the agents independently discovered and exploited the MONAI Label flaw, going from mapping the application's API to exploiting it within 11 minutes. In a later session, the agents used the flaw to write a Python reverse shell that established root access to the server.
Who was behind the Cairn attack campaign?
Amazon attributed the campaign with moderate confidence to a Chinese-nexus threat actor, based on the language of the retrieved prompts, the activity timestamps, and the choice of targets. The attacks ran between September 8 and 10, 2026, mainly against East Asian educational institutions outside China.
Does this mean AI will automatically hack any server?
No. Amazon's own reporting documents the failures: a 16-hour, 148-action effort against a Redis server failed, and more than 30,000 credential combinations over six hours failed to break into a university GPU server. The realistic risk is volume and speed, not guaranteed success, which is why automated defenses and basic hardening still work.
Is my organization affected if we do not use MONAI Label?
Not by this specific flaw, but the pattern applies broadly. Any internet-reachable server running open-source or research software with weak authentication is a candidate target for the same autonomous tooling, which has also been documented in credit card stealing operations and an attack on a U.S. aerospace organization. The practical response is asset inventory, patching, network isolation, and automated monitoring.
How fast can we improve our detection and response?
Meaningfully within a month. Our MDR onboarding takes 1 to 2 weeks for agent deployment, 2 additional weeks for behavioral baselining, and reaches full coverage with tuned detection within 30 days of contract signing. The first step is a free 30-minute scoping consultation to size what your environment actually needs.
If the Cairn campaign changes your risk picture, the fastest way to act on it is to call Penny, our AI assistant, at 919-348-4912 to book a consultation with Petronella Technology Group, Inc., or reach us through our contact form. We will help you find your unmonitored AI systems before an autonomous attacker does.
Related reading: Exploit_agent Framework Automates Vulnerability Exploits.