Safety2 min read

Hackers Use Stolen AI Keys to Fund Their Own AI Attacks

By , Senior AI ConsultantPublished

Free AI hacking tools borrow paid AI models to run, so criminals now steal other people's AI keys, and any small business that builds on AI is in the sweep.

Hacking software that lets AI agents plan and carry out break-ins with little human help has moved from security testing into everyday crime. One such tool, Artex, has taken some intrusions from first target to administrator access in under three hours, and it is not alone: Cairn, CyberStrikeAI and PentAGI are being used the same way.

The attack pays for itself

Artex has no intelligence of its own. It hands each goal to an AI model such as Claude or GPT and turns the answers into commands for its agents, so every attack uses up paid AI time. A criminal who wants to go after hundreds of targets at once needs a lot of it, and the cheapest source is someone else's account.

That is what one large campaign did. Records that Amazon's decoy servers captured from the agents show them searching the internet for AI apps that hold keys: tools that sit between a company and its AI providers, and builders that let people without programming skills make chatbots. Each key they found was tested, and the working ones were fed into the framework's own gateway, where they paid for the agents hunting for more keys. The agents graded every find by whether it worked and what it could reach, and they treated keys that had been public for more than 30 days as low value.

That last rule shows how fast this runs. The attackers assume that a key leaked a month ago has already been noticed and cancelled, so the danger is in the first days after a leak.

You do not have to be chosen

For example, a shop owner uses a no-code builder to put an AI assistant on the shop's website, and leaves the administrator password at its default. An agent finds the page, takes the provider key stored inside and tests it. The owner never sees an attack. They see a strange AI bill, or a locked account, at the end of the month.

The attacker did not pick that shop. It was one line in a sweep, and running the sweep over ten thousand more lines costs almost nothing extra. This is the reason small businesses are now in range: the work that used to make a small target not worth the effort is now done by agents, and several targets run at the same time.

The weak spots the tool looks for are the ones nobody owns: partner portals, staff-only apps, old services and the connections between systems. Every company has a few. The forgotten login page is the usual way in.

What closes the door

The takedown route that worked, slowly, for Cobalt Strike is closed here, so the defence has to sit with the people who hold the keys. Four habits cover most of it.

Keep AI keys out of code, public repositories and chat messages, and store them in a proper secrets tool. Set a monthly spending limit on every key at the AI provider, so a theft has a ceiling. Change the default password on any AI builder or gateway you run, and keep its admin page off the open internet. And if a key has ever been public, cancel it at once and make a new one; waiting to see whether anyone uses it is the mistake the attackers are counting on.

Share this

STAY INFORMED

Get AI intelligence like this delivered to your inbox.

Free forever · Unsubscribe anytime


You May Also Find Valuable