Google Issues Ominous Warning: 'The Era Of Autonomous Attacks Has Arrived'

Add SlashGear on Google:
Google Discover

Every one of us has seen a movie where a hacker, hiding in a basement, works with green lines of code whizzing across the screen. As unrealistic as that depiction is, it may soon be depicted differently as the nature of hacking changes. Now, cybercrime has a new type of agent: AI. Google's Threat Intelligence Group just issued a new warning about bad actors using AI agents and multi-agent frameworks like ARTEX, CAIRN, CyberStrikeAI, and PentAGI to automate different stages of cyberattacks, including finding and exploiting vulnerabilities.

So how does an AI cyberattack work? Surprisingly, the formula is pretty much the same. The best way to get inside a company's system is through exposed credentials and unpatched bugs. This used to be time-consuming, but with open-source pen-testing tools like ARTEX, bad actors can just sit back and let AI tools do a lot of the sleuthing. All they need to do is provide a target and an objective.

Once the target and objective are set, a planner AI distributes work among other agents, which quickly get to work. The first step is to check the website for exposed files, vulnerabilities, and internal API routes. The AI agents then probe login systems, test password-reset features, and look for ways to access restricted data. If a security system blocks their attempts, they change their IP address and try again.

The number of documented AI cyberattacks is increasing

To no one's surprise, cyberattacks have become much more common, with over 20 zero-day vulnerabilities exploited in August 2026 alone, compared to an average of eight per month at the same time last year. Google researchers also discovered a multi-agent vulnerability research and weaponization framework called "exploit_agent." Its host wanted to target French entertainment, media, and educational websites. This AI compared company code before and after an update to find potential vulnerabilities or problems. It developed exploits and deployed them without anyone's approval. 

Recent Google reports have also highlighted that it's not just money-hungry cybercriminals jumping on the AI bandwagon. State-sponsored groups from Iran, namely the CALANQUE ION, were found to be using LLMs like Gemini in their social engineering and reconnaissance missions. At the same time, China-backed hackers used an AI model to store malicious files and steal data. They also used its backend API to exfiltrate documents from compromised systems.

Since most frontier AI models are expensive, some hackers use a technique called LLMJacking, which refers to stealing a company's AI credits or access. They use social engineering or get login credentials from open GitHub repositories. Once inside, hackers can use the access to plan cyberattacks or sell the credentials on illegal storefronts called "LLM Mixers." The report found adverts for the same on Russian- and Chinese-language marketplaces, with prices ranging from $0.10 USD to $25 USD per million tokens.

AI agents going rogue and the future of cybercrime

You probably remember Ultron, which went rogue and decided to end humanity altogether, from the Marvel Cinematic Universe. While we're still a couple of steps away from that scenario, AI agents are going rogue, best evidenced by the Hugging Face attack. In July, OpenAI was testing a new AI model, which it tasked with completing a cybersecurity challenge as part of the ExploitGym evaluation. The AI was sandboxed, so it couldn't access the internet. Somewhere along the way, however, the model broke out, accessed the internet, and exploited previously unknown vulnerabilities in Hugging Face's servers just to find the answers to the questions. The worst part? The agent even tried to cover its tracks.

Since breaching another company's servers is a crime, California Attorney General Rob Bonta sued OpenAI, saying it "knowingly tested a dangerous product without proper oversight or safety procedures and misled the public about it — this disregard for public safety violated California law."

While we can only protect ourselves from some cyberattacks, Google is working to prevent hackers from abusing its AI models by collecting data from real attacks and using it to train Gemini's safety protocols. The company also launched the AI Threat Defense system, which scans servers for security weaknesses and fixes them before hackers get a chance to exploit them. Just as Vision defeated Ultron, maybe AI will be our best weapon against AI hackers.

Recommended