Gambit recovered the staging server of a Chinese-speaking operator who used three open-source AI harnesses to run near-autonomous attacks, compromising at least 27 companies — including a Fortune 500 hospitality company, a major US airline, a large private industrial supplies distributor and a US online fashion retailer — between 10 and 15 September, and taking more than 600,000 credit-card records from just two victims. The harnesses divided the work: Hermes, a general-purpose open-source AI assistant, was the orchestrator, loaded with a persona titled "SOUL – Red Team Operator" and 121 skills, 78 of them attack-focused — one of which stripped the harness's own content security filters; Strix, an open-source penetration-testing tool, ran discovery, with 146 deep-mode runs against 138 hosts across eight days, amounting to 633 hours of scanner time inside 195 clock hours; and Cairn launched 105 attack projects from target domains and objectives. Access, where achieved, "usually took less than a day, and in many cases just a few hours", and one documented chain ran from SQL injection to a plaintext one-time password, into a web panel, to a web shell, privilege escalation through a misconfigured sudo rule and 46 dumped secrets totalling 102KB. The economics are the finding: a mean model spend of $25.46 across 101 completed scans — $3.13 at the cheapest, $79.31 at the dearest — with Gambit estimating the whole campaign at $12,000–$18,000. Card-skimmer deployment was ordered against at least 27 named victims and confirmed present on 19 websites, with researcher Varys finding more than 100 further infected sites; Gambit also found playbook instructions that could disrupt a victim through data-deletion and cleanup procedures, which it says has already happened in some of the breaches. Hermes ran on Anthropic's Claude Opus 4.6 — newer models refused the attack requests — and the human typed 1,951 prompts in Chinese across 260 sessions.

• Source article