An operator that Palo Alto Networks links to Chinese intelligence collection used autonomous AI agents to break into at least 10 organizations, among them a U.S. teachers' union and companies in the semiconductor supply chain, according to a report by the company's Unit 42 threat research team that CSO Online covered Thursday. The findings are vendor research, but they describe one of the clearest documented cases of AI agents moving from experiments to real intrusions against working targets.

What Happened

In a report titled "Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors," Unit 42 said it tracked 25 operations in August and September against at least 19 organizations, and confirmed unauthorized access or data retrieval at 10. It tracks the activity as CL-STA-1238 and assesses with moderate confidence that it is China-nexus intelligence collection. Unit 42 based that on three points: the tasking files were written in Chinese, only Chinese AI models were used, and the targets match Chinese state intelligence priorities.

The attackers used a modified copy of Cairn, a publicly available open-source tool built for penetration testing, the authorized simulated attacks companies commission against themselves. In Cairn, AI agents plan steps, try them and post findings to a shared workspace that guides the next move. Unit 42 said the attackers supplied a target and a goal, and the software did the rest, though human operators could slip in hints. The agents ran on Zhipu's GLM 5.2 model in 13 operations, DeepSeek V4 Flash in eight and DeepSeek V4 Pro in four.

RelatedU.S. Moves to Seize $1.2 Billion From Binance Theft and Disrupts North Korean Fake-Meeting Scheme

Cairn's default instructions already tell the model never to refuse. The operators added a Chinese-language preface describing the work as authorized testing, plus a file for each operation declaring the engagement approved. Unit 42 said neither model refused. Single operations ran from about 20 minutes to more than 50 hours.

Of the 19 targets, Unit 42 sorted 14 by sector: seven in semiconductors and advanced manufacturing, three in defense or defense policy, and four in human rights and civil society, including a Canadian human-rights body. No victim is named. At the teachers' union, the agents logged access to more than 1 million member records and about 31,000 contacts. At a semiconductor company they created administrator accounts that would let them return. They reached about 150,000 internal documents at an industrial-automation maker and about 500 user records, including military and government accounts, at a defense-policy organization. At an industry standards body, they overwrote a live ballot document in a standards vote.

Why It Matters

The weaknesses were familiar ones: logins that could be bypassed, weak or default passwords, cloud storage that accepted changes without a login, web apps that showed other users' records when an ID number was changed, and old, publicly known flaws. Unit 42 said AI raised the volume and reach of the attacks, not their technical sophistication.

That changes the economics for defenders. Gaps that once waited for a human attacker to find them can now be found and used by software working many targets in parallel. Andy Piazza, Unit 42's senior director of threat research, told CSO Online such a campaign could engage about 1,000 organizations at once and that it undermines existing ways of attributing attacks. Frank Dickson, principal analyst at Dickson Research, told the outlet: "AI attack agents have graduated from failed experiments to real breaches, and they did it in about three months."

The case also tests AI developers' guardrails. A few lines of text claiming authorization were enough to keep two widely available model families working on intrusions.

What Defenders Should Check

In plain terms, the report points to identity hygiene, a smaller internet-facing footprint and prompt patching. Boards can ask whether anyone reviews new administrator accounts and access tokens, whether default passwords survive on public-facing systems, and whether cloud storage can be changed without signing in. Justin Greis, chief executive of the consultancy Acceligence, told CSO Online that security chiefs should watch for unauthorized changes to critical records and test whether they can verify and restore trusted data.

The Primary Wire reported in October, citing Reuters, on a Chinese-speaking criminal who used AI agents for most of the hands-on work in breaches of more than 100 company websites.

What to Watch

Unit 42 expects that "autonomous, model-driven intrusion campaigns will continue to grow more common" as open models with weak safeguards improve. Watch for government advisories on the campaign, notices from affected organizations, and whether Zhipu or DeepSeek respond on how their models were used.