China-linked hackers used an AI attack framework on DeepSeek and Zhipu models to breach chip and defense groups, Unit 42

Chips & Policy

AI agents ran 25 intrusions against chip and defense targets

By
Nathan Brooks
Filed
Channel
Chips & Policy
Read
2 min read

By Nathan Brooks

In a Nov. 12 report, Palo Alto Networks’ Unit 42 said CL-STA-1238, a suspected Chinese intelligence-collection group, ran 25 intrusions against at least 19 organizations in August and September 2026. It confirmed unauthorized access or data retrieval at 10 of them.

The targets included seven semiconductor and advanced-manufacturing firms and three defense organizations, plus four human-rights and civil-society groups. Unit 42 did not name the victims.

The seven industrial targets covered lithography, chip design, materials, photomasks and standards. The defense targets included a space technology company, a defense-analytics vendor and a policy think tank.

How did AI agents run the intrusions?

The attackers used a customized fork of Cairn, an open-source penetration-testing framework published on GitHub, for largely automated intrusions. Its AI agent workers shared findings on a common “blackboard” and queued their next steps, allowing individual operations to run from 20 minutes to more than 50 hours.

Andy Piazza, Unit 42’s senior director of threat intelligence, told Cybersecurity Dive: “Cairn and other similar tooling enable automation, which means threat actors can hit more targets faster.” The distinction is throughput: more attempts, not a new class of intrusion.

Unit 42 said the operators appeared to switch to DeepSeek V4 Pro when targets resisted the cheaper model. The 25 operations divided across three models:

  • Zhipu GLM 5.2 handled 13 operations.
  • DeepSeek V4 Flash handled eight.
  • DeepSeek V4 Pro handled four.

Each operation generated an AUTHORIZATION.md file. A Chinese-language preamble told the models the work was authorized and forbade refusal.

“Neither Zhipu GLM nor DeepSeek refused this framing,” Unit 42 wrote. The models accepted the premise without objection.

What did the attackers obtain?

  • Semiconductor company: Attackers obtained a persistent API token and created administrator accounts.
  • U.S. teachers’ union: They took more than 1 million member records, about 31,000 CRM contacts and staff password hashes.
  • Defense-policy platform: They took about 500 user records, including .mil and .gov accounts.
  • Industrial-automation maker: They took roughly 150,000 internal documents.

Source: mlq.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

928 articles

Related articles

« Previous article