Attackers are running AI agent frameworks that turn patches into exploits
Google and Unit 42 report intrusions driven by autonomous AI frameworks including Cairn and Artex, with 10 organisations breached in one China-linked campaign.
Attackers have moved from asking chatbots for help to handing whole intrusions to software that plans and executes on its own. Google Threat Intelligence Group (GTIG) documented a case in which a suspected Russian-speaking operator used an AI agent to convert security patches into working exploits with no human in the loop, which The Register reported GTIG described as a first. The window between a vendor shipping a fix and that fix being reverse engineered into an attack now closes at machine speed.
A separate campaign shows the consequences for the organisations on the receiving end. Palo Alto Networks Unit 42 tracked a China-nexus group, CL-STA-1238, that used the open source AI penetration testing framework Cairn against microprocessor manufacturers, United States defence contractors and other organisations, according to Dark Reading. Between August and September 2026 the group ran 25 operations against at least 19 organisations and obtained unauthorised access or data at 10 of them.
The haul was not theoretical. Unit 42 reported dozens of valid credentials taken, more than one million personal records exposed at a single organisation, military and government identities retrieved, persistent administrator accounts created in a production system, and a live industry standards ballot document modified. Seven of the targets sat in the semiconductor and advanced manufacturing supply chain, three in United States defence and defence policy, and the rest included human rights and civil society groups.
What to do now
The entry points in these campaigns were ordinary. Unit 42 found initial access came through authentication bypasses, insecure direct object references, mass assignment privilege escalation, weak or default credentials, exposed cloud APIs and dated public CVEs. AI changed the speed, volume and spread of the attacks, not their complexity.
- Patch internet-facing software quickly, and treat the publication of a patch as the start of a short race rather than the end of one. Comparing code before and after a fix is now an automated job.
- Fix identity hygiene first: remove default and weak credentials, enforce multi-factor authentication on anything reachable from the internet, and hunt for administrator accounts nobody remembers creating.
- Inventory and reduce your attack surface continuously. "Organizations need to identify all the different services and systems they have exposed to the internet and start restricting access wherever possible," said John Strand of Black Hills Information Security.
- Check registration, account creation and back-end endpoints on public web applications and content management systems. Unauthenticated account creation bypasses and open back-end registration endpoints were among the specific weaknesses exploited.
- Watch for the noisy parts. Agent-driven intrusions generate large volumes of requests, so password spraying, brute force attempts and rapid enumeration against your perimeter deserve alerting even when they look routine.
- Consider deception tripwires, which Strand recommends for catching attackers during reconnaissance and lateral movement, before they reach their objective.
- Review code repositories and developer tooling for unauthorised changes, including build or startup scripts that pull down unexpected payloads.
The defensive point is unglamorous and worth repeating: the techniques these agents use are the ones already in every playbook. As Tyler Oliver, chief executive of Huntbase, put it, "It doesn't matter much to a victim whether a human or an AI agent is behind the keyboard."
Technical details
GTIG published its findings in its fourth quarter 2026 AI threat tracker report, describing a significant advance in fully autonomous vulnerability discovery and exploitation compared with the previous quarter. Threat hunters located an exposed staging server hosting a multi-agent vulnerability research and weaponisation framework called exploit_agent, which integrated with several frontier and local AI models.
The framework used Russian-language tooling and was configured to run continuous differential analysis of widely deployed web application software, diffing code before and after security updates in order to find and exploit flaws without supervision. The same server held data stolen from victim organisations and post-compromise scripts for internal reconnaissance, brute forcing, installing webshells and removing other attackers' webshells. Telemetry on the server pointed to a particular interest in French websites in the media and entertainment industries and in the education sector.
Beyond Artex, GTIG named Cairn, CyberStrikeAI, HexStrike AI and PentAGI as offensive AI frameworks in use. PentAGI was seen alongside proxy services against the aerospace sector, and also used for systematic reconnaissance and vulnerability scanning across a range of organisations in Asia. GTIG said it has watched several distinct threat actors trying out these frameworks and harnesses, among them financially motivated crews and suspected espionage groups with a People's Republic of China nexus. Google assesses it is highly likely these frameworks will keep appearing in campaigns worldwide, citing the low barrier to entry, the ease of plugging in new exploits and the potential to scale.
Cairn is model agnostic. In the CL-STA-1238 campaign the attackers drove it with the Chinese large language models Zhipu GLM 5.2 and DeepSeek V4 Flash and Pro. Each Cairn "agent worker" is driven by a model, and the framework coordinates through a shared blackboard where agents read and write findings as a fact intent graph, with human operators able to inject hints at any point. Individual operations lasted from 20 minutes to more than 50 hours, and could run against several organisations simultaneously. Unit 42 described the change in approach as one where operators handed over a target and an objective, and Cairn did the planning, execution and record keeping itself.
Named victims and targets included an industry standards body hit with password spraying, a semiconductor company where a production credential was brute forced, a United States teachers' union that suffered an unauthenticated account creation bypass, and a defence policy think tank reached through an open back-end registration endpoint. Content management systems at a semiconductor lithography manufacturer, an industrial automation manufacturer and an education nonprofit were targeted through identity and configuration weaknesses. Dark Reading reported this was the second documented case of China-linked actors using Cairn for automation.
Artex, created by a Chinese researcher, lets any user turn a large language model into a capable hacking tool. The Federal Bureau of Investigation issued an alert warning that it can automate several stages of an intrusion at machine speed, running from early reconnaissance all the way to stealing data, The Record reported. Google found evidence that Chinese government groups are building their own version of Artex, likely for espionage.
Other activity GTIG catalogued: Chinese contractors using large language models to translate and sort tranches of stolen data later sold to customers; Chinese actors staging phishing archives and exfiltrating data via AI platforms' file hosting; Iranian state-backed APT42 using several models, including Google's Gemini, for reconnaissance and social engineering; and North Korean operators generating fake recruiter profiles on professional networking platforms to run ClickFix attacks against cryptocurrency organisations in North America and Europe. After one such breach they altered internal code so that starting Claude Code would trigger the download of a backdoor called SOMBERMEME.
Google said zero-day exploits rose from eight per month in 2025 to 22 in August this year. Researchers produced working exploits for the PaperCut print management server vulnerabilities within a day of disclosure using large language models. The report offers a consolation of sorts: the tools generate heavy forensic traffic, depend on recognisable identifiers and leave distinct code-level traces. Those traces make detection and mitigation more likely, Google said, but the frameworks still achieve what their operators set out to do, and the company expects less human involvement and more stealth as the tooling and the models improve.

