Gambit Security: AI agents hack retailers for $25.46 a target

Gambit Security's investigation began with an unusual asset: a relay server belonging to one of the attackers. From that vantage point the firm could watch how the campaign actually ran, and what it saw was not a human crew typing commands but a chain of open-source tools dividing labor. Strix scouted targets, Cairn pursued entry for hours on its own, and Hermes coordinated the job using a library of 121 skills. Human input was sparse — 1,951 prompts across 260 sessions, only a few instructions per target.

What makes the campaign distinctive is that the AI did not repeat one fixed attack. Cairn assessed each target's website and server state and built an attack path on the spot, so the vulnerabilities exploited varied widely by victim. In one documented case, a SQL injection flaw led to a database, one-time passwords were read to defeat multi-factor authentication, an admin panel was breached, a file-upload weakness allowed code execution, and the trail eventually reached a Magento database where encrypted credit card numbers could be decrypted. Targets were not random either: the attackers pulled shopping sites from access-ranking services, excluded those on major EC platforms, and prioritized stores running custom-built systems likely to carry unknown flaws.

The economics are the part Gambit Security calls more important than the attack's scale. The main tools are open source, and adding one more target costs only tens of dollars, while AI can work in parallel without tiring. Companies that once seemed too small to be worth attacking may now fall inside the attackers' reach — a shift that hinges on how far these low costs fall and how quickly defenders can respond to self-restoring malicious code. Gambit Security cautions that its report is a provisional finding, and with many attack projects still unexamined, the true scale and damage may prove larger than the numbers published so far.