Bitget: haz trading inteligente. Lionel Messi, Official Partner. Regístrate

Google’s threat intelligence group says attackers are already using language model systems that analyze software updates, deduce which flaw was fixed and build an exploit without human intervention. Google is responding with Google AI Threat Defense, which combines Gemini and other models, Wiz, CodeMender’s automatic patching and Mandiant. John Hultquist, its chief analyst, sums it up: attackers are going after known, unpatched flaws, “but with much more speed and scale”.

The company’s threat intelligence group detected systems capable of searching for vulnerabilities, preparing exploits and executing several phases of an intrusion. Automation speeds up operations, although it can still leave traces that help detect them.

Cybercriminals are moving from asking artificial intelligence models for occasional help to integrating agents capable of chaining tasks within an operation. That is one of the main conclusions of the latest report from the Google Threat Intelligence Group, published on November 16. The team, together with Mandiant, describes cases in which tools based on large language models were used to probe systems, search for vulnerabilities and advance intrusions with less human intervention. (Google Threat Intelligence report)

The shift is not that AI has invented a new way to break into networks. According to the report, attackers continue to exploit stolen credentials and unpatched systems, two avenues known for years. The difference lies in speed and scale. Agents can automate tasks that previously required manual work and respond to what they find during the operation.

Google says it has observed several multi-agent frameworks used in real intrusions. Among them are open-source tools such as CAIRN and PentAGI, as well as tests with other platforms. These systems can combine security utilities, generate code for a specific task and adjust their actions based on the results obtained.

One of the examples described in the report is a system called exploit_agent. It was configured to compare program versions before and after a patch was released, locate potential flaws, and prepare and execute exploits. Google says the process could run without a person intervening at each step. The observed activity targeted French media, entertainment and education sites.

Automation also shortens the time between the release of a patch and the appearance of exploitation attempts. Attackers can compare the fixed code with the previous version to deduce which vulnerability has been repaired and develop an attack against systems that have not yet been updated. At the same time, the ease of producing proofs of concept with AI is generating more code that looks valid but does not work, adding verification work for security teams.

The report also documents a market around so-called LLMJacking. In these cases, attackers steal cloud credentials or computing resources to use AI models or resell access to third parties. Google and Mandiant describe incidents in which fraudulent use of these resources sent billing soaring within hours. In the most serious cases cited, charges for a single day exceeded $100,000 and the cumulative abuse reached $1 million before it could be stopped.

AI also appears in influence operations. Google maintains that state-linked groups, commercial providers and networks seeking revenue from political content are already incorporating it to produce synthetic text and media, create fake profiles and distribute posts at scale. The report warns, however, that more convincing content does not necessarily mean the campaign achieves its goals.

The ability to operate quickly comes at a cost for attackers. According to Google, some autonomous tools leave easily traceable technical activity, generate conspicuous traffic or use recognizable identifiers. That lack of stealth can make it easier for defenders to detect and contain an intrusion. The company expects this weakness to diminish as the models and the systems that use them improve.

For organizations, the practical message is familiar, although the window to react is narrowing. Keeping systems updated, protecting credentials and monitoring the use of cloud services remain essential. What is new is that an exposed vulnerability or a compromised account can be exploited more quickly and through increasingly automated processes.

Get the best of the week in your inbox

Every Monday: the most-read stories on Territorio Blockchain, the new podcast episode and news about our events. Free, no spam.