Breaking into an online shop used to take skill and time. A campaign uncovered by security firm Gambit shows how much AI has changed that. One attacker, typing only a few instructions per target, used AI agents to compromise at least 27 online retailers in five days, steal more than 600,000 unexpired credit cards and plant card skimmers on other stores.
The cost per target averaged about $36 in AI fees. For every small retailer that assumed it was too minor to be worth a hacker’s time, that number changes the maths.
How the AI agents ran the attack
Gambit recovered the attacker’s staging server and rebuilt the campaign from its logs, it explained in its published report. Three open source AI tools did most of the work: Strix searched sites for weaknesses, Cairn chained together the break in, and Hermes coordinated the jobs and remembered what worked.
The agents ran on models bought through the OpenRouter marketplace, including DeepSeek and GLM, and on an older Anthropic model after newer ones refused the requests. The human operator typed 1,951 prompts across 260 sessions, often brief instructions in Chinese such as reading a vulnerability report and starting work.

What it cost and who was hit
Between 10 and 15 September the attacker launched 105 attack projects. Across 101 completed scans, the average cost per target was about $36, ranging from roughly $4 to $113. Over four weeks the recorded spend came to almost $10,000. Gambit estimates the full seven week campaign cost between $17,000 and $26,000.
Speed was the other shock. The median time to compromise a store was under a day, and some fell within hours. The skimmers were hidden in several ways, from extra JavaScript and Google Tag Manager tags to poisoned page caches and scheduled server jobs.
Victims included a Fortune 500 hospitality company, an airline and stores selling fashion, bicycles, beauty products, wine and firearms. Gambit confirmed skimmers on five stores and detected more than 100 other infected sites. One victim was a retailer in New Zealand, and the attacker favoured smaller shops running custom code. At one bicycle retailer, the agent’s clean up routine deleted 180 database tables, including the store’s own backups.
Why Australian retailers should care
Most of the stolen cards were issued in the US, but the targeting was global and nothing in it depends on geography. A small Australian store on an ageing platform with a few unpatched plugins is exactly the kind of target these agents handle in hours. A breach involving customer card data is also likely to trigger obligations under the Notifiable Data Breaches scheme.
The basics matter more than ever: patch the platform, remove unused plugins and admin accounts, watch for unexpected scripts on checkout pages, and keep credentials out of code and shared documents. A business password manager is one of the cheapest fixes. For a wider review of how AI changes your risk, AI consulting should include your own exposure, not just your use of AI.
Plan for recovery, not just defence
Gambit’s sharpest advice is about recovery. It warns that data loss can arrive as a side effect of someone else’s clean up routine, and that a plan ending at “the database is restored” does not answer what the business needs to keep running. Agents that work around obstacles are not only a criminal tool, as the OpenAI agent in the Medicare portal showed. Every retailer should know which systems it cannot trade without and prove it can bring them back.
