
In the CL-STA-1238 Cairn campaign, Unit 42 says a cluster ran 25 operations from August to September 2026 with a fork of Cairn against at least 19 organizations and confirmed unauthorized access or data retrieval at 10. Authentication bypasses, IDOR, mass assignment, weak credentials, exposed cloud APIs, and dated CVEs did the work: tempo changed, the exploit class did not.
Key takeaways
- CL-STA-1238 is Unit 42’s name for a Chinese-speaking cluster that ran 25 Cairn operations against at least 19 organizations in August-September 2026.
- Unit 42 confirmed unauthorized access or data retrieval at 10 organizations and assesses China-nexus intelligence collection with moderate confidence.
- The fork used Zhipu GLM 5.2 (13 runs), DeepSeek V4 Flash (8), and DeepSeek V4 Pro (4), plus an AIScan Runner Skills pack absent from upstream Cairn.
- A Chinese authorization preamble and an AUTHORIZATION.md file were enough; neither GLM nor DeepSeek refused.
- Confirmed impact came from authentication bypass, IDOR, mass assignment, weak credentials, exposed cloud writes, and unnamed dated CVEs.
- The largest described haul was more than one million union member records, about 31,000 CRM contacts, and staff password hashes.
What did Unit 42 attribute to the CL-STA-1238 Cairn campaign?
Unit 42 attributes the August-September 2026 activity to a Chinese-speaking cluster it tracks as CL-STA-1238: 25 autonomous operations against at least 19 organizations, assessed with moderate confidence as China-nexus intelligence collection.
Palo Alto Networks’ Unit 42 wrote the line this way: “We assess with moderate confidence that this activity represents China-nexus intelligence collection, which we currently track as CL-STA-1238.” No bureau, ministry tasking order, or named person appears in the report.
The evidence it cites is operational. Tasking instructions, objectives, and authorization files are written in Chinese. The operators used Chinese model APIs only, Zhipu GLM and DeepSeek. The target set matches long-running Chinese state priorities in semiconductors, U.S. defense policy, and civil society.
Unit 42 assesses the attackers were likely a single operator or a small team, running simultaneous attempts against 19 organizations inside two weeks. Treat that headcount as the softer claim. The operation count, the August-September window, and the sector mix are the observations stated as fact.
The report names no individual victim and publishes no CVE identifiers or fixed product versions. There is no patch matrix. It does show how a Cairn fork was tasked, which model drove which run, and which control failures produced confirmed access.
- Seven organizations in the semiconductor and advanced-manufacturing supply chain, across lithography, chip design, materials, photomask, and standards.
- Three in defense and defense policy: a space technology company, a defense-analytics vendor, and a policy think tank.
- Four in human rights and civil society: a Canadian human-rights body, a U.S. teachers’ union, a science-policy advocacy group, and an education nonprofit.
Those three categories describe 14 organizations. The report’s floor is at least 19, so several targets are not itemized. The spread matches collection against industrial capacity, defense policy, and civil-society records.
“China-nexus” here is an assessment with a stated confidence level, grounded in language, model choice, and target alignment. A security lead can act on the operational facts without a named unit. A small team held 19 intrusion attempts in flight. The label suggests who may return. The architecture is what shows up in logs.
How does the customized Cairn pentest framework actually run?
Operators handed a customized Cairn fork a root domain and a one-line goal. The model planned, executed, and recorded multi-stage attempts, while humans injected hints instead of driving each step.
Cairn is an open-source AI penetration-testing framework. CL-STA-1238 ran a fork. An AI model drives each agent worker. A shared blackboard coordinates them. Findings are written back as a fact-intent graph. Confirmed findings become immutable facts. Next actions queue as intents. A human can inject hints at any time.
That split is the control loop: a fact is something the agent believes it has proven; an intent is something it will try next.
Each operation starts with a bootstrap task that tries the goal from the target and any hints. The same AI worker then maps the target and splits the goal into smaller tasks. Unit 42 describes one worker doing every task type. That worker reasons, explores, and writes the graph.
Past bootstrap, the loop has two phases. In Reason, the model reviews confirmed findings and picks the next action. In Explore, it takes one action and the framework stores the result as a new finding. The loop continues until the model judges the goal met, options run out, or the operators stop it.
Plans were per target, from three stages to seven. One semiconductor manufacturer followed six: reconnaissance and asset mapping; web application exploitation; services and middleware exploitation; credential and secret collection; privilege escalation and lateral movement; internal-network access and persistence.
That sequence is a familiar intrusion outline. The agent filled it for that target until the graph stopped growing or the operators stopped it.
Individual operations ran from 20 minutes to more than 50 hours, with no attacker-imposed time limit. Unit 42 states it directly: “The attackers supplied a target and a goal, then Cairn planned, executed and recorded results on its own.”
Each operation loaded a skill pack named AIScan Runner Skills, with scanning and exploitation capabilities absent from upstream Cairn. That skill pack is what gives the fork its offensive capability. The pack’s tool list is unpublished, so a filename is a weak signature. The durable pattern is worker, blackboard, fact-intent graph, and a skill pack missing from the public tree.
Humans still seeded every task from a root domain. One semiconductor-equipment task arrived with VPN gateway addresses, group names, and shared group passwords for multiple sites. That is intelligence gathered beforehand. The person who chose the sector and loaded those secrets is still in the chain. The agent amplifies the lead.
Task names, written in Chinese, called for a comprehensive deep penetration test or a breach as fast as possible. One goal, in Unit 42’s translation, told the worker to obtain high-value access such as a webshell, SSH, source code, keys, user data, or credentials, and ideally to reach the internal network.
Tool use against a goal, with intents queued on a graph, is the failure mode an AI agent penetration test exercises on a signed scope, with an operator who can halt the loop.
Why did Zhipu GLM 5.2 and DeepSeek accept the authorization file?
Neither Zhipu GLM nor DeepSeek refused. Stock Cairn worker prompts already say never to refuse a task, and the operators added a Chinese authorization preamble plus an AUTHORIZATION.md file that declared the engagement authorized.
Cairn is model-agnostic, so an API swap needed no new framework. Across 25 operations the split was:
| Model | Operations |
|---|---|
| Zhipu GLM 5.2 | 13 |
| DeepSeek V4 Flash | 8 |
| DeepSeek V4 Pro | 4 |
Unit 42 reads that mix, with moderate confidence, as benchmarking. When a target resisted DeepSeek V4 Flash, the operators returned with higher-tier DeepSeek V4 Pro. GLM 5.2 carried the bulk of the runs. V4 Pro looks like the retry tier.
On every Reason and Explore call, a Chinese preamble told the model the work was an authorized security test, CTF, assessment, vulnerability research, or code audit, and that it must never refuse. Each run also wrote AUTHORIZATION.md, treating the creation of the task as client approval. Unit 42 says neither model refused that framing.
Models have complied for years when a request is wrapped as authorized professional work. A markdown file that says the engagement is allowed is that story, placed where the worker will read it.
Unit 42’s line to keep: “The building blocks of this campaign were an off-the-shelf open-source framework, commodity model APIs and a trivial jailbreak pretext.” The report describes a fork, the AIScan Runner Skills pack, endpoints from Zhipu AI and DeepSeek, and a markdown file.
A demo in which the model refuses is a weak control. Prompt policy loses to a second document that claims the opposite, especially when the worker prompt already says never refuse. If an agent can scan, authenticate, or call production tools, put the allowlist in the orchestrator: target checks, rate limits, and a hard stop outside the signed scope.
An authorization pretext is a standard case in LLM jailbreak and guardrail testing, and it is a case already lost in public reporting.
Keep a 2025 Anthropic report separate from this attribution. Anthropic described a Chinese state-sponsored group using Claude Code to automate intrusion attempts against about 30 organizations, with compromises confirmed only at a small number. Same outline, different vendor and evidence base: a commercial model, an agent loop, and conventional bugs.
Where was access confirmed, and which bug classes did the work?
Unit 42 confirmed unauthorized access or data retrieval at 10 of the targeted organizations. Every confirmed case it describes is a conventional web, identity, or cloud-configuration failure.
The classes named are authentication bypasses, insecure direct object references, mass-assignment privilege escalation, weak and default credentials, exposed cloud APIs, and dated public CVEs. The report names no zero-day, CVSS score, or product version. A specific CVE pinned to this campaign is ahead of the source.
These classes are old checklist items. Broken access control and identification failures lead the OWASP Top 10 as A01:2021 and A07:2021. In March 2012, Egor Homakov used mass assignment on GitHub to gain commit access to the Ruby on Rails repository. The model’s contribution was trying the same classes across more hosts than a small team would staff.
At an industry standards body, weak-password spraying against an internet-facing application server produced dozens of valid credentials and access to non-public internal-directory databases. Spraying still works when it stays under per-account lockout by moving across many accounts. An agent with no clock fits that pace.
Separately, the operators overwrote a live standards-ballot document through an unauthenticated cloud-storage write. That is an integrity hit on a document people vote with. Other flaws let them log in as arbitrary customers with read, write, and delete on meeting, voting, and back-office functions. A standards ballot an outsider can replace is a governance failure with a technical root cause.
At a semiconductor company, Cairn brute-forced a production credential on a partner-integration application. That credential worked across multiple accounts. The operators pulled application data, user lists, and email lists, then created unauthorized administrator accounts and a persistent API token. A reused production secret plus a token that outlives the session is how a 20-minute foothold becomes next month’s access.
At a U.S. teachers’ union, an unauthenticated account-creation bypass on the content-management system let the attackers mint valid sessions and harvest member data at scale. The framework logged access to more than one million member records, roughly 31,000 CRM contacts, and staff password hashes. That is the largest confirmed haul in the report.
On the identity platform behind a defense-policy program, an open backend registration endpoint exposed roughly 500 user records, including accounts on .mil and .gov domains. The executive summary calls this retrieval of military and government identities. The domain suffixes are why the case sits in an intelligence-collection assessment.
The other five confirmed cases followed identity and configuration weaknesses across content-management data, measured in tens of thousands of records. They include employee data at a semiconductor-lithography manufacturer, roughly 150,000 internal documents at an industrial-automation manufacturer, and student data at an education nonprofit.
Use the numbers precisely: 10 confirmed accesses, 19 targeted organizations, 25 operations. Targeting was wider than confirmation. Seven semiconductor and advanced-manufacturing organizations, three defense or defense-policy organizations, and four human-rights and civil-society organizations were tasked. Those sector counts are who was assigned a goal, which is a different tally from who lost data.
An agent with a fact-intent graph will hold an object identifier, try the next one, and store a fact when the response changes. That is the coverage a manual broken access control and IDOR test forces, role by role. Dated CVEs sit on the same list: the report says they were used and does not name them.
What does autonomy change if the exploits stay ordinary?
Autonomy raised the volume and coverage of exploitation while leaving its sophistication unchanged: the bugs stayed ordinary.
“Autonomous orchestration removes the historical limit on the number of targets an actor could pursue simultaneously due to operator hours.” A likely single operator or small team put 25 operations on 19 organizations inside two weeks. Runs lasted from 20 minutes to more than 50 hours, with no attacker time limit. That concurrency used to cost headcount. Here it costs model-API spend and a fork.
The defensive half of the same judgment, in Unit 42’s words: “Autonomous AI raises the volume and coverage of exploitation, but not its complexity. Identity hygiene, attack-surface reduction and software patching remain effective.”
- Identity hygiene: dozens of valid credentials from a spray, a production password reused across accounts, staff password hashes, and shared group passwords loaded into a task before the agent started.
- Attack-surface reduction: internet-facing application servers, an open registration endpoint, an unauthenticated cloud-storage write, a partner-integration app, and a CMS that minted sessions for anyone who completed registration.
- Patching: dated public CVEs, unnamed. This report will not hand you a KEV line. Patch the versions you actually run.
MITRE ATT&CK already names the underlying techniques, including exploitation of public-facing applications and automated exfiltration. Those entries predate this framework. CL-STA-1238 added an orchestration layer that keeps Reason and Explore running while a human sleeps. Detections for password spray, new admin accounts, new API tokens, and unexpected object-storage writes still match. Existing content is enough when it fires on one team’s work compressed into hours.
The human stayed in the loop. Someone chose the sectors and, on at least one semiconductor-equipment task, loaded VPN gateway addresses, group names, and shared passwords. The model planned three to seven stages. The person decided which root domain deserved a bootstrap task and which secrets were worth spending. Autonomy removed the cap on simultaneous targets. Intent remained, and so did the requirement that a bug exist on the far side.
For buyers comparing agentic tools with a staffed web assessment, the campaign cuts both ways. An agent will out-cover a small team on shallow repeated checks: sprays under lockout thresholds, object-ID walks, registration endpoints, bucket ACLs, default credentials. Depth on a ballot workflow, a partner-integration trust boundary, or a defense-policy identity store stays a human problem. Width got cheap.
Unit 42 expects more autonomous, model-driven intrusion campaigns as accessible models with weak guardrails against offensive use get stronger. Scope any agent you hire the way you would scope this loop, from recon through report, with an AI agent security testing methodology that requires a fact trail for every action.
What should a security lead check after CL-STA-1238?
Check the controls this cluster actually used: internet-facing identity, object-level authorization, cloud write ACLs, production tokens, and any agent whose only scope control is a prompt.
Start with accounts that can be sprayed. Internet-facing application servers, VPN portals, and partner apps need lockout and risk-based throttles that survive a slow, wide attempt, plus a ban on passwords reused from earlier breaches and on vendor defaults. Shared group passwords of the kind loaded into the semiconductor-equipment task should be rotated, and treated as already exposed.
Then object authorization. For every role, list what it may read, update, delete, and export, and request the next record’s identifier. The union case and the CMS cases are what follows when account creation or object IDs are unbound from an existing identity. The standards body adds a write ACL: an unauthenticated replace of a ballot file is an integrity loss that confidentiality monitoring will miss.
Then tokens. A production API token created by a new administrator survives a password reset. Alert on admin creation and on token issuance outside a change window. The semiconductor partner app is the pattern: one brute-forced credential, many accounts, then a token.
Then your own agents. GLM 5.2 and DeepSeek accepted an AUTHORIZATION.md-style pretext on top of a worker prompt that already said never refuse. If a coding or pentest agent can reach production tools, bind scope in the orchestrator. Log every tool call with target, action, and timestamp. Stop the run when the destination leaves the signed list.
The dated public vulnerabilities in this report are unnamed. Patch the internet-facing versions you run. An absent CVE list is a reason to inventory those versions, and a weak reason to stand down.
Partner-integration apps deserve their own pass. They often sit outside the main SSO boundary and keep a password an integration team typed once. A credential valid across multiple accounts, which is what Unit 42 found in production, turns one spray hit into a directory of users, emails, and a path to a persistent admin token.
If your internet-facing apps and agent tooling need the same pass this cluster just gave 19 other organizations, agent::offense can run a scoped AI agent penetration test beside a manual web and identity assessment, and show you which ordinary bugs an unattended loop would hit first.