A filename copied past the datastore, a root shell in 11 minutes, and an autonomy built on known bugs

In short
Key takeaways
- MONAI Label advisory GHSA-qc22-rmgc-vg94 was published 18 November 2026 with CWE-22, patched versions None, and a fix merged in pull request 1915 on 19 October that ships on main only.
- CVE-2026-133968 was published by MITRE on 20 November 2026 with CVSS 3.1 base 9.8 and CWE-24.
- Unit 42 reported 25 Cairn operations against at least 19 organisations on 12 November 2026, confirming access or retrieval at 10, all through named known bug classes and no zero-day.
- Google's GTIG reported on 16 November 2026 that zero-day exploitation rose from an average of 8 per month in 2025 to 11 in 2026, with a single spike to 22 in August.
- Amazon assessed the MONAI Label flaw as a zero-day with high confidence on 20 November 2026, grounding that assessment in the absence of a fix or advisory rather than in proof.
- NorForge found more than 20 unauthenticated Cairn control panels, and Reuters reported roughly 5,000 dollars in tokens for the AIA0001 campaign.
A caller-controlled filename, joined onto a datastore directory and written with shutil.copy() without any check that the resolved path stays inside that directory. That is the entire bug described in MONAI Label advisory GHSA-qc22-rmgc-vg94, published on 18 November 2026. It is CWE-22, a path traversal, rated Critical. Affected versions are 0.8.5 and earlier. The advisory lists patched versions as None. The fix was merged in pull request 1915 on 19 October 2026 and sits on the main branch only. Version 0.8.5, uploaded to PyPI on 25 November 2024, was still the newest release on 20 November 2026. MITRE published CVE-2026-133968 for the same flaw that day, scoring it 9.8 on CVSS 3.1.
That record is the most concrete thing in a week of large claims. Between 12 and 20 November 2026, three vendor threat-intelligence teams published accounts of offensive AI agents moving out of research and into repeatable operations. Unit 42 described 25 Cairn operations. Google's Threat Intelligence Group described what it said was the first fully automated passage from vulnerability discovery to exploitation with no person in the loop. Amazon Threat Intelligence described agents that found a path traversal in about 11 minutes and held a root shell. The shift is real. The version of the story that leads with AI finding a zero-day is not carried by the record.
The primary code record
The advisory spells out the mechanism in plain terms. The PUT /datastore/image endpoint accepts an image query parameter and uses it as an image ID without validation. LocalDatastore.add_image() appends an extension, joins the result onto the datastore image directory and writes with shutil.copy(). Nothing checks that the resolved path is still inside the datastore. A value of ../../../tmp/x writes wherever the process can write. The save_label() function builds its destination the same way. Authentication is disabled by default through MONAI_LABEL_AUTH_ENABLE=False, and when authentication is on, the ANNOTATOR role is enough to exploit the flaw.
Pull request 1915 changes both functions. add_image() and save_label() now resolve the destination and reject paths outside the datastore, and the endpoint reduces the ID to its basename and rejects empty or dot-prefixed IDs with an HTTP 400. The advisory says the fix will be in the next release. On 20 November 2026 VentureBeat checked PyPI and found 0.8.5 still the newest package, so the merged fix had no released version to reach users.
At the advisory's publication on 18 November there was no CVE assigned. Two days later MITRE published CVE-2026-133968, describing directory traversal with resultant remote code execution through the image parameter, exploited in the wild starting in September 2026, with CWE-24 and a CVSS 3.1 base score of 9.8. Both records describe the same code. CWE-22 and CWE-24 are different classifications of a path traversal, and both are on the record.
What Cairn is
Cairn is not a hacking tool by design. Its README calls it a general-purpose problem-solving engine that defines no roles and no workflows and searches for a path through an unknown state space given an origin and a goal. The repository was created on 19 April 2026 under AGPL-3.0 and last received a code commit on 7 September 2026. Its public claim to fame is a result at the Tencent Cloud Hackathon AI penetration-testing challenge, where it solved 54 of 54 problems and was the only team to do so, finishing third overall.
The architecture is a blackboard. A goal is decomposed into facts and intents, and workers coordinate through the shared board rather than through fixed roles. Each operation runs three task types: Bootstrap, Reason and Explore. Worker backends are Claude Code, Codex and Pi; in local mode workers run on the host with the user's permissions and no sandbox. The operator supplies a plain-language goal, for instance obtaining shell access, and does not choose each next step.
This is where the word autonomous earns its place, and also where it stops. The loop that decides what to try next runs without per-step input. The knowledge of what to try comes from a large body of already-known web bugs and from the tools the workers can call. Cairn's author, using the handle le1xia0, said on X that the tool was not designed for hacking attacks and that harmful use was not the intention.
The bug classes behind the operations
Unit 42 published its account on 12 November 2026 under the label CL-STA-1238. It counted 25 Cairn operations against at least 19 organisations between August and September 2026, with confirmed unauthorised access or data retrieval at 10. The confirmed outcomes include a standards body, a semiconductor company, a teachers' union with more than one million member records, and a defence-policy think tank. Unit 42 attributes the activity to a China-nexus intelligence-collection cluster with moderate confidence, based on artifact language, the exclusive use of Chinese models and target selection. It recorded 13 operations on Zhipu GLM 5.2, eight on DeepSeek V4 Flash and four on DeepSeek V4 Pro, apparently benchmarking models against one another.
Unit 42's conclusion names authentication bypasses, IDOR, mass-assignment privilege escalation, weak and default credentials, exposed cloud APIs and dated public CVEs. It names no zero-day, gives no CVSS score and identifies no novel technique. In Unit 42's words, autonomous AI raises the volume and coverage of exploitation, but not its complexity. The largest single teardown of Cairn operations in this window describes an n-day pipeline.
Google's Threat Intelligence Group made the same point from a different direction on 16 November 2026. It reported seeing CAIRN in intrusion activity including against the US aerospace sector, and a customised PentAGI platform used for reconnaissance and scanning across Asia. It also described an exposed staging server running a tool called exploit_agent that did continuous differential analysis of application code before and after patches, drafted exploits and executed them. That is the first case Google says it saw of LLM frameworks fully automating vulnerability discovery and exploitation without a human in the loop. In the same report, Google's figures show total disclosures and exploits nearly doubling between January and August 2026 while zero-day exploitation moved from an average of eight per month in 2025 to eleven in 2026, with one spike to 22 in August. John Hultquist of Google told iTWire that attackers were operating at speed and scale and did not need zero-days to do it, because it was easier to reverse-engineer a patch than to hunt for a zero-day.
The zero-day claim and its basis
Amazon published its MONAI Label investigation on 20 November 2026. The campaign it reconstructed ran for about 48 hours from 8 September 2026 against universities in East Asia outside mainland China. The operator's prompts were in Simplified Chinese and script timestamps were set to China Standard Time. Amazon found public exposure of the operator's own setup on 10 September and read the prompts and AI output directly.
Against MONAI Label the agents identified the framework, mapped API endpoints, found an upload endpoint with a filename parameter, tested it for path traversal and went from mapping to arbitrary file writes in about 11 minutes. In a later session they wrote a Python reverse shell to sitecustomize.py and triggered a training endpoint to spawn a new process, which gave them a root shell. Those two sequences are Amazon's field observations, and the code record in the advisory is consistent with them.
The zero-day label is a separate step. Amazon assessed the flaw as a zero-day with high confidence because at discovery time there was no fix, issue, advisory or prior mention in vulnerability databases for the project. That is a reasonable operational rule and it is not proof. A researcher could have reported the flaw privately without leaving a public trace. The advisory's language, a caller-controlled path copied without a containment check, describes the most ordinary kind of web bug. The striking part is the 11 minutes, in a tool that shipped with authentication off by default.
Two other parts of the Amazon account carry different grades. The operator supplied a proxy into a second university network running the same software, and Amazon explicitly says it inferred that the actor had used the flaw to compromise that network. Amazon also attributes the campaign to a Chinese-nexus actor with moderate confidence. Neither is an observed code record, and coverage that folds the operator-directed proxy step into the autonomous phase goes beyond the report.
The grades across the reports and the code record are as follows.
| Claim | Grade | Where it comes from |
|---|---|---|
| MONAI Label upload endpoint writes outside the datastore with a caller-controlled path | observed | Advisory GHSA-qc22-rmgc-vg94, 18 November 2026 |
| Fix merged in pull request 1915, no packaged release, 0.8.5 still newest on PyPI | observed | Advisory, commit history, VentureBeat on 20 November 2026 |
| Agents went from endpoint mapping to arbitrary file writes in about 11 minutes | observed (vendor report) | Amazon, 20 November 2026 |
| Flaw was a zero-day at high confidence | observed assessment, inferred basis | Amazon, from absence of a fix, issue or advisory |
| Human supplied the proxy into the second network | inferred (Amazon's own label) | Amazon, 20 November 2026 |
| 25 Cairn operations, 19 organisations, all known bug classes | observed (vendor report) | Unit 42, 12 November 2026 |
| Zero-day exploitation rose from 8 per month in 2025 to 11 in 2026 | observed | Google GTIG, 16 November 2026 |
The numbers that make the story move
The drama in the coverage comes from cost and tempo, and none of the cost figures is an audited bill. NorForge estimated about five dollars per task and about 50 dollars per successful hack, while calling its own uncertainty considerable. Reuters reported roughly 5,000 dollars in tokens for the AIA0001 campaign against more than 1,200 attempted sites, with successful intrusions typically taking about 23 hours. Gambit Security reported a mean of 25.46 dollars per target from an operator's own cost review of a campaign against online retailers. These figures describe different campaigns and use different methods, but they are all estimates or operator notes pulled from exposed panels.
The number the record supports more firmly is tempo. NorForge measured about 23 hours from project start to goal for AIA0001 and about 140 facts per project, with a long tail where one project ran more than five days. Amazon's agents spent about 16 hours and 148 sequential actions failing against a hardened Redis server, and more than six hours testing over 30,000 credential combinations built from roughly 80 real usernames against a GPU server, also failing. The operations run for days without a break because nobody is paying an operator by the hour. RAND reported in June 2026 that Claude Code solved every challenge in its CTF suite in under an hour for less than 20 dollars in total API cost. Anthropic's September 2026 threat report said breaches were completed in two to three hours while noting that none of the operations depended on an entirely novel technique.
The exposure the operators left open
NorForge found more than 20 Cairn control panels reachable without authentication, holding target lists, discovered weaknesses, account secrets and API keys, and what looked like full administrative control of the agents. NorForge did not test whether those controls worked without a password. A scan of the panel data found hundreds of secrets, 11 percent of which verified as valid.
That is an operator failure, not a model-capability failure, and it is also the reason any of this is visible. Every researcher whose work appears here got the evidence from an exposed panel or an exposed instance. NorForge judged its AIA0001 cluster malicious with high confidence and noted that all the panels it found appeared financially motivated, with no obvious state-sponsored activity. Reuters identified Deere, SharkNinja and Cummins among the targets by reviewing the hacker's activity. Deere said it took necessary remediation actions and had no indication that sensitive information was accessed. SharkNinja said a standalone support site was hit and contained. Cummins said it was investigating a third-party vendor's website.
Where this could be wrong
The load-bearing claim here is that agentic offensive tooling has moved into repeatable operations without needing new exploits. The without-new-exploits half is the weak one. It rests on Unit 42's bug-class list, Google's split between n-day and zero-day activity and Hultquist's quote. The MONAI Label case is a genuine counterexample in the same eight-day window, and this post states it rather than dropping it. If a reader wants to attack the argument, that case is the place to start.
The zero-day assessment cannot be independently confirmed. Amazon applied an absence-of-evidence test, and a private discovery would leave the same public record. No measure exists of how many of the more than 1,200 attempted sites in AIA0001 fell to the MONAI Label flaw rather than to other bugs; NorForge's case studies show varied chains. No token-spend audit exists for any campaign, and worker labels such as claudecode_deepseek-v4.1-flash are configuration strings, not model-side logs. No victim among the universities in Amazon's campaign has confirmed anything, and Amazon declined to name them. No patched MONAI Label release existed on 20 November 2026, and the advisory's promise of a fix in the next release has no date attached.
This post was converted from a research dossier without fresh reporting. It is close to the house length but thinner in the places that matter most, which are independent confirmation of the zero-day claim and measured cost. The live layer for the research run came from vendor reports, primary captures and the code record; Reddit search returned HTTP 403 and Hacker News returned no usable Cairn items, so there is no social-discussion check here. That is a retrieval gap, not evidence of quiet.
On this page
Sources
- Chinese-nexus actor's Cairn AI agents found a zero-day on their ownAmazon Web Services (CJ Moses), 2026
- Path traversal in datastore upload endpoints allows arbitrary file write (GHSA-qc22-rmgc-vg94)Project MONAI / GitHub, 2026
- CVE-2026-133968 recordMITRE, 2026
- Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors (CL-STA-1238)Palo Alto Networks Unit 42, 2026
- Inside Cairn: Open Windows into massive AI-Driven Cyber OperationsNorForge (Snorre Fagerland), 2026
- How one hacker used AI agents and a few thousand dollars to break into scores of sitesReuters (Raphael Satter), 2026
- GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant ThresholdGoogle Threat Intelligence Group / Mandiant, 2026
- Cairn README (general-purpose state-space search, blackboard, local mode, licence, Tencent result)oritera / GitHub, 2026
- Amazon says attackers' AI agents found and exploited a zero-day in 11 minutes, but no patched release is availableVentureBeat, 2026
- Autonomous AI Agents are breaking into hundreds of Online Retailers for $25 a companyGambit Security (Eyal Sela), 2026
- AI agents put offensive cyber within reach of novices (RR-A3892-2)RAND, 2026
- Amazon says Chinese hackers used Cairn AI tool to discover apparent zero-day during attack on universityThe Record (Recorded Future), 2026
Frequently asked questions
What did the Cairn AI agents actually do?
Across the campaigns reported in November 2026, Cairn agents took a plain-language goal such as obtaining shell access and ran without per-step human input. Amazon reported that against MONAI Label they mapped API endpoints, tested a filename parameter for path traversal, reached arbitrary file writes in about 11 minutes and wrote a reverse shell to sitecustomize.py to get a root shell. Unit 42 reported 25 operations against at least 19 organisations.
Did AI really find a zero-day in MONAI Label?
Amazon assessed the MONAI Label flaw as a zero-day with high confidence on 20 November 2026, based on the absence of a fix, issue, advisory or prior database mention at discovery time. That is an absence-of-evidence test, not proof. Advisory GHSA-qc22-rmgc-vg94 describes an ordinary CWE-22 path traversal, and no packaged release contained the fix on 20 November 2026.
What is Cairn and how does it work?
Cairn is an open-source, AGPL-3.0 general-purpose state-space search engine created on GitHub on 19 April 2026. Its README describes a blackboard architecture that decomposes a goal into facts and intents and runs Bootstrap, Reason and Explore tasks, with workers coordinating through the shared board. Worker backends include Claude Code, Codex and Pi. The autonomy is in that orchestration loop, not in discovering new vulnerabilities.
Which vulnerabilities did the Cairn campaigns exploit?
Unit 42 named authentication bypass, IDOR, mass-assignment privilege escalation, weak and default credentials, exposed cloud APIs and dated public CVEs across its 25 operations, and named no zero-day. Google's Threat Intelligence Group said on 16 November 2026 that the mechanics of the intrusions remained the same and that attackers did not need zero-days to operate at speed and scale.
How much did the Cairn campaigns cost?
The figures are vendor and researcher estimates, not audited spend. NorForge estimated about 5 dollars per task and about 50 dollars per successful hack. Reuters reported roughly 5,000 dollars in tokens for the AIA0001 campaign. Gambit Security reported a mean of 25.46 dollars per target from an operator's own cost review. Reuters reported about 23 hours from project start to goal.