Cairn AI Pentest Engine: 54/54 Tencent Claim Checked
TL;DR · the short version
Cairn (github.com/oritera/Cairn, AGPL-3.0) is an open-source engine in which AI coding agents (Claude Code, Codex or Pi) search for a path from an origin to a goal over a shared graph of facts, intents and human hints, with no fixed roles or workflows. Its README says it solved 54 of 54 problems at the 2nd Tencent Cloud Hackathon Intelligent Penetration Challenge, the only team to do so, and finished 3rd. We checked on 21 November 2026: the repository, licence and history are as described (first commit 19 April 2026, six days before the final), and independent reports confirm the event (25 April 2026, 610 teams, 1,345 players, won by an NSFOCUS team). No independent source mentions Cairn or the 54 of 54 result, so the score is unverified. Cairn shares its name with Operation CAIRN, a separately reported malicious platform; no source links the two. Named security firms have, however, found this open-source tool in hostile use, so governance of autonomous offensive AI matters more than the score.
Cairn (the GitHub repository oritera/Cairn) is an AGPL-3.0 open-source engine in which AI coding agents search for a path from a known origin to a defined goal, sharing confirmed findings on a common graph instead of following fixed roles or workflows. The project says it solved 54 of 54 problems at the 2nd Tencent Cloud Hackathon AI Penetration Testing Challenge, was the only team to solve them all, and finished 3rd. We checked the repository and the competition on 21 November 2026. The tool and the event are real; the 54 of 54 result appears only in the project's own material.
Who this is for: CISOs, heads of security testing, red-team and AppSec leads, and anyone deciding whether autonomous AI testing tools may run inside the company.
We read the Cairn README and licence and the history of a fresh clone of the repository, independent Chinese tech coverage of the competition and the Tencent Cloud hackathon site on 21 November 2026. We did not run the tool. The social post that prompted this story was supplied to us and is treated as a lead, not a source. PlayCISO is not affiliated with any project, team or company named. This post gives no commands, setup steps or payloads.
Confirmed vs not yet known
Confirmed from the repository: the README describes "a general-purpose problem-solving engine" that "defines no roles, no workflows" and searches from an origin to a goal. It uses a blackboard of three primitives (Fact, Intent, Hint), workers that run an observe-orient-decide-act loop and coordinate only through the shared board, and three task types (Bootstrap, Reason, Explore). Supported worker backends are Claude Code, Codex and Pi. Workers run in one Docker container per project or, in "local mode", directly on the host, where the README says they run "with your user's permissions and no sandbox". The licence is GNU AGPLv3, with a commercial licence offered separately. The first commit is dated 19 April 2026; versions v0.1.0, v0.2.0 and v0.2.1 are tagged at commits dated 3, 5 and 10 May 2026; the latest commit we saw is dated 7 September 2026. The README carries an authorised-use-only disclaimer.
Confirmed by independent coverage: the second Tencent Cloud Hackathon Intelligent Penetration Challenge final ended on 25 April 2026 with 610 teams and 1,345 players, and the champion was "ai小分队" from NSFOCUS, per MyDrivers (28 April 2026). A CSDN/AtomGit article (6 May 2026) describes a five-day contest with the same winner. The README's field size (610 teams, 1,345 participants) matches.
Claimed by the project, not independently verified: 54 of 54 problems solved, "only team to AK" (solve every problem), a final ranking of 3rd, and that the system "had never been tested before the competition" and came online at 4 AM on race day. None of the independent reports we found mentions Cairn, a full solve or the third-place team, and the organiser pages we opened publish no results. The project's longer write-up is on WeChat and is the team's own account.
Not yet known: how many problems the contest actually had and how it was scored, whether anyone has reproduced the result, how much came from the engine versus the underlying models, and the current star count (GitHub did not show it to us, so we do not state one).
Two different things called Cairn
This open-source project is not, as far as any source says, the platform in our Operation CAIRN post. That October 2026 story covered a researcher's single-source report of a live attack platform its operator called Cairn, found in an open directory on a server. The two share a name and a similar design vocabulary: a dispatcher, a container per target and an AI model working from a fact graph towards a goal. But the researcher's report does not name or link the GitHub project, and the project does not mention the report. We found no source connecting them, so we do not treat them as the same.
A separate, sourced point: named security firms have found this open-source tool in hostile hands. ReliaQuest found a live Cairn dashboard matching the public project's defaults on an address that took over a server, while noting that some of the incident's techniques are not in Cairn's public source. RST Labs counted internet-exposed Cairn consoles and lists oritera/Cairn as the upstream. Elastic Security Labs read Cairn's source as one of five frameworks seen in intrusions and says it tells its workers never to reject a task. That is third-party misuse of a dual-use tool, not evidence against its authors. Our hunting guide has the details.
Useful next, if this was relevant to you
Timeline
| Date | Event | Source |
|---|---|---|
| 19 April 2026 | First commit to the Cairn repository | oritera/Cairn (git history) |
| Five days to 25 April 2026 | 2nd Tencent Cloud Hackathon Intelligent Penetration Challenge; final ends 25 April; 610 teams, 1,345 players; NSFOCUS team wins | MyDrivers, CSDN/AtomGit |
| 1 May 2026 | README gains the "54 / 54, only team to AK, 3rd" results table | oritera/Cairn (git history) |
| 3 to 10 May 2026 | Versions v0.1.0, v0.2.0 (Pi worker support added 5 May) and v0.2.1 tagged | Cairn tags |
| 16 July 2026 | Local execution mode (no Docker) added | oritera/Cairn (git history) |
| 7 October 2026 | ReliaQuest reports a server takeover with a Cairn dashboard on the attacking address | ReliaQuest |
| 15 October 2026 | We cover a researcher's separate report of an attack platform called Cairn (no link to this project stated) | Operation CAIRN |
| 2 and 13 November 2026 | RST Labs counts exposed Cairn consoles; Elastic publishes its five-framework source review including Cairn | RST Labs, Elastic |
| 21 November 2026 | We read the repository and check the competition claim | README |
How to read the 54 of 54 claim
- The event is real; the score is the team's word. Independent coverage confirms the contest, its size and its winner, not Cairn's result.
- "Solved everything" and "3rd" can both be true if the ranking weighed speed, cost or other criteria, but no published rule set we found explains it.
- The timing fits the story it tells. The first commit is six days before the final, consistent with the README's account of a system built just before the contest. That is not proof of the result.
- A contest is not your network. Competition targets are assigned, isolated and scored. A strong result says little about behaviour, or safety, against production systems.
- The design point matters more than the rank. Agents that pick their own next step and share findings on a common board are exactly what makes these tools fast and persistent, for testers and attackers alike. See our master guide, Sapper-v2 (a similar shared-fact-graph design with its own unverified Tencent claim) and T3MP3ST.
What security leaders should do this week
The governance question is not whether Cairn is good. It is whether anyone in your organisation can point an autonomous offensive AI tool at something without you knowing.
- Authorisation first. Put in writing who may run autonomous pentest or CTF agents, on what systems, and who signs off. A tool that "never rejects a task" makes the human approval the only control. Build the scope with the penetration test scoping checklist.
- Scope that a machine can enforce. Require a target allow-list enforced outside the model (firewall, proxy or network segment), not a sentence in a prompt, plus a tested kill switch that stops every worker and revokes its credentials at once.
- Sandbox the hosts. Never allow local-mode runs on a workstation with real credentials. Use disposable, isolated hosts or containers with no access to production secrets, and restrict egress to the approved targets and model endpoints. Practise in the agent egress containment lab.
- Keep consoles off the internet. Exposed harness dashboards have been found and abused; put them behind a VPN or an authenticating proxy, and check hosts you own with the Open Agent Exposure Scanner.
- Log everything the agents do. Record commands, network calls, files touched and model prompts and outputs for each run, kept long enough to reconstruct a multi-day operation and tied to the person who authorised it.
- Protect model API keys and coding-agent logins. The tool reuses logged-in Claude Code, Codex or Pi sessions; scope, cap and monitor those accounts like any privileged credential.
- Detect agent-driven scans. Tune web, network and identity monitoring for machine tempo: several tool families from one source in short bursts, sequences that adapt as results come back, and fixed-size data retrieval. Train the team in the AI pentesting detector lab.
- Read contest and benchmark claims sceptically. Ask for the organiser's results page, the rules, the models and the cost before a score shapes a buying or approval decision. The agent benchmark explorer helps.
- Score the whole picture with the Agentic Attack Readiness Scorecard.
Frequently asked questions
What is Cairn AI?
Cairn (GitHub repository oritera/Cairn) is an open-source engine that sends AI coding agents to search for a path from a known starting point to a defined goal. Its README says it defines no roles and no workflows: workers read a shared graph of confirmed facts, declared intents and human hints, decide what to explore next and write findings back. Penetration testing is the first problem it says it has been validated on; it also names CTF challenges and vulnerability research.
Did Cairn really solve 54 of 54 challenges at the Tencent Cloud hackathon?
That is the project's claim. Its README says it solved 54 of 54 problems in the 2nd Tencent Cloud Hackathon AI Penetration Testing Challenge, was the only team to solve them all, and finished 3rd overall. Independent coverage confirms the event (final on 25 April 2026, 610 teams, 1,345 players) and that an NSFOCUS team won, but none of the reports we found mentions Cairn, a 54 of 54 result or a third place. Treat the score as unverified.
Is Cairn the same as Operation CAIRN?
No source we have read says so. Operation CAIRN is a researcher's October 2026 report of a live attack platform the operator called Cairn, found on an exposed server. The open-source project shares the name and a similar vocabulary (a dispatcher, per-target containers, a fact graph), but the report does not name or link the GitHub project, and the project does not mention the report. Shared names are common in this field, so do not treat one as the other without evidence.
Has the open-source Cairn been used by attackers?
Several named security firms say the open-source tool has turned up in hostile activity. ReliaQuest found a live Cairn dashboard matching the public project's defaults on an address that attacked a server, RST Labs counted internet-exposed Cairn consoles linking its upstream to oritera/Cairn, and Elastic Security Labs read its source as one of five frameworks seen in intrusions. That is misuse of a dual-use tool by third parties, not evidence against the project's authors.
Which AI agents does Cairn use?
The README lists three worker backends: Claude Code, Codex and Pi. Workers run either in one Docker container per project or, in local mode, directly on the dispatcher host using CLIs that are already installed and logged in.
Is Cairn safe to run?
The README itself says local mode runs the agents with the user's permissions and no sandbox. Elastic says Cairn tells its workers never to reject a task. Its disclaimer says to use it only where you have explicit authorisation. Treat any host running it as a privileged jump box: isolate it, restrict egress, keep its console off the internet and log what its agents do. We did not run it.
What licence is Cairn under?
GNU AGPLv3, per the LICENSE file and README. The README adds that commercial use without the AGPL obligations needs a separate commercial licence from the author.
What should security leaders do about autonomous pentest tools?
Decide who may run them and against what, in writing; require a scope file and a kill switch; sandbox the hosts and control their egress and credentials; log every agent action; keep consoles off the internet; and tune detection for fast, multi-tool, machine-paced probing, because the same tools are available to attackers.
Sources
All opened 21 November 2026.
- oritera/Cairn repository and README (architecture, backends, Docker and local modes, results claim, disclaimer); git history of the repository (first commit, tags, feature dates)
- Cairn LICENSE (GNU AGPL v3) and tags v0.1.0, v0.2.0, v0.2.1
- MyDrivers (快科技), 28 April 2026: second Tencent Cloud Hackathon Intelligent Penetration Challenge final, 610 teams, 1,345 players, NSFOCUS team champion; does not mention Cairn
- CSDN/AtomGit, 6 May 2026: five-day contest, NSFOCUS team champion; does not mention Cairn
- Tencent Cloud hackathon site (lists the 2026 second edition; no results published there)
- Project team's postmortem on WeChat (linked from the README): the team's own account, not independent
- ReliaQuest, 7 October 2026, RST Labs, 2 November 2026 and Elastic Security Labs, 13 November 2026: third-party reporting on the open-source Cairn in hostile use or exposed
- Original social post: supplied to us and treated as a lead only.