
Hackers, who used Russian in technical documentation identified by Google, set up a series of AI agents to “continuously compare pre- and post-patch code, identify vulnerabilities, and automatically develop exploits without human intervention,” Google said.
That means the agents independently found software flaws and ways to use them. The attackers targeted French websites in the media, entertainment and education sectors, Google said.
Until now, even the most advanced AI models could not carry out cyberattacks entirely independently; they often required skilled hackers to guide them or fix mistakes. The key change instead was that AI models allow hackers to accelerate attacks and make them more targeted.
The development is likely to spook governments and critical infrastructure operators who are still reeling from this summer's hack by an OpenAI agent into the Hugging Face platform. Though acting autonomously, that hack didn't appear to be maliciously aimed at certain targets like the one being reported by Google. Fears are also running high after the recent discovery of major flaws in PHP, a programming language that underpins 70 percent of the web, and a series of hacks on supermarket fridges, which have been partially blamed on Iran.
This article has been updated.