• Sources: primary, discussion
  • Summary: Unit 42 reports that from August to September 2026 an actor it tracks as CL-STA-1238 ran a customized fork of the Cairn AI penetration testing framework in 25 operations against at least 19 organizations in semiconductor manufacturing, defense policy and civil society. The attackers gave each task a one-line goal and a root domain, ran 13 operations on Zhipu GLM 5.2 and 12 on DeepSeek V4 Flash or Pro, and added a preamble claiming the work was authorized, which neither model refused. Cairn confirmed access or data retrieval at 10 organizations, including more than one million member records at a US teachers' union, persistent admin accounts and an API token in a semiconductor company's production system, and an overwritten standards-ballot document.
  • Why it matters: The flaws exploited were weak passwords, auth bypasses, IDORs, mass assignment and exposed cloud APIs, so identity hygiene and attack surface reduction now face machine-speed probing at many targets at once.
  • Background: Agentic pentest frameworks pair a language model with a loop that plans, runs tools and records results, so an operator only supplies a target and a goal. Cairn coordinates its model-driven workers through a shared blackboard where confirmed findings accumulate as facts and queued next steps as intents. The digest reported on 2026-11-13 that Cloudflare counted 50 such AI pentest frameworks scanning its customers, and this report is a documented case of one used for intrusion rather than testing.