AI agents steal 600,000 card records and plant skimmers on 119 sites

September 28, 2026 at 08:30 PM2 min

This digest was compiled by AI from multiple sources — links to the originals are below.

Security researchers at Gambit uncovered an AI-driven campaign that stole data on more than 600,000 payment cards and injected web skimmers into at least 119 online stores since July 2026. The attacker used open-source tools Strix, Cairn, and Hermes to automate scanning, exploitation, and post-exploitation. Gambit analysts accessed the attacker's staging server and found direct evidence of the attacks.

Key Facts

The campaign stole data on more than 600,000 active payment cards from two victims.
At least 119 websites were infected with web skimmers.
The attacker spent $7,005.71 on OpenRouter over four weeks, with total AI costs estimated at $12,000–$18,000.
From August 23 to 31, Strix ran 146 times against 138 hosts, totaling 633 hours.
Between September 10 and 15, the attacker conducted 105 attack waves and compromised at least 27 organizations.

Attack Infrastructure

The campaign relied on three open-source tools: Strix for scanning and vulnerability discovery, Cairn for exploitation and gaining shell or admin access, and Hermes for coordination, post-exploitation, and tactical decisions using claude-opus-4.6. The operator only provided brief goals to the AI agents, which then performed most of the work autonomously. From August 23 to 31, Strix was launched 146 times against 138 hosts, totaling 633 hours of operation. Between September 10 and 15, the attacker conducted 105 separate attack waves and compromised at least 27 organizations.

Skimmer Deployment Methods

The method of injecting the web skimmer depended on the level of access and system architecture. In various cases, the attacker added malicious code to legitimate JavaScript files and payment pages, embedded it in Google Tag blocks, modified S3 and CDN content, server cache, and database fields. Sometimes the attacker changed Kubernetes deployment configurations and created cron jobs that restored the web skimmer after removal. Compromised organizations included a Fortune 500 hotel company, a major US airline, an industrial equipment supplier, and an online clothing store.

Evidence and Costs

Gambit analysts accessed the attacker's staging server and found direct evidence of the attacks. After stealing card data, the AI agent deleted it from Magento databases, as instructed in one of Hermes' skill files, causing data loss and disruptions for several affected stores. The AI agents allowed the attacker to scale attacks to hundreds of targets at low cost. Over four weeks, the campaign's OpenRouter account spent $7,005.71, and total AI expenses for the studied period were estimated at $12,000–$18,000. The attacker's own calculations showed the average cost of scanning one target was $25.46, ranging from $3.13 to $79.31.