Latest

AI-generated illustration of a robotic arm holding a key beside a cracked padlock. It is not a picture of anything in this story. Illustration: Grok.

By news editor

Attackers are now running software that finds and exploits vulnerabilities with no person involved, and intruders who hijack corporate AI accounts have consumed as much as $1 million in resources before being cut off, according to Google's threat intelligence unit.

The findings appear in the latest AI Threat Tracker from Google Threat Intelligence Group, known as GTIG, written with the company's Mandiant incident response unit. It names no victim organizations and gives no total for the intrusions that used the tools it describes.

On automation, the report is specific. GTIG writes: "In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time." LLM stands for large language model, the kind of AI system behind chatbots and coding assistants.

Its evidence is an exposed staging server hosting a framework the report calls exploit_agent. By Google's description, the framework compares the code of popular web application software before and after a patch is announced, identifies potential flaws, drafts exploits and runs them without human intervention. Data on the server suggests the operator was interested in French websites tied to the media and entertainment industries and the education sector, the report says.

For companies, the pressure falls on patching schedules. GTIG says attackers may find it more efficient to build exploits for flaws that already have a fix, called n-day vulnerabilities, because the patch narrows what they need to study. The report describes the effect as a shrinking patch window: less time for a business to install an update before the flaw is attacked.

Overall vulnerability disclosures and exploits nearly doubled between January and August 2026, GTIG says. Zero-day exploits, which hit flaws before any fix exists, rose more slowly, from an average of 8 per month in 2025 to 11 per month in 2026, before reaching 22 in August.

Separately, GTIG says it has observed two open source penetration-testing frameworks, CAIRN and PentAGI, deployed in real intrusions, and has seen experimentation with two more. Such frameworks chain AI agents to carry out stages of an attack. CAIRN was used against targets including in the US aerospace sector, and a platform built on PentAGI scanned a wide range of organizations in Asia between late August and early October 2026, according to the report. Those experimenting with the tools include financially motivated groups and suspected espionage groups with a China nexus.

The direct losses come from what the report calls LLMJacking: breaking into a company's cloud environment to use its computing capacity and paid AI services. Mandiant responded to several such incidents during 2026, the report says, without giving a number. On cost, it states: "In some cases, single-day fraudulent charges exceeded $100,000 USD. In the most severe cases, cumulative resource abuse reached $1 million USD before remediation." The intrusions it describes affected financial services, retail, transportation and technology organizations in North America, South America and the Asia Pacific region.

Stolen access is then resold. GTIG describes storefronts it calls LLM Mixers that route buyers' requests through pools of stolen credentials, at advertised prices ranging from $0.10 to $25 per million tokens, the units in which AI usage is metered.

The report also describes limits on the attackers. The agentic frameworks, it says, "currently trade stealth for velocity" because unattended runs generate traffic and leave code traces that raise the odds of detection. LLMs can invent details or even targets, which makes operations less reliable. And after patches for two PHP vulnerabilities were released on Oct. 21, 2026, hundreds of claimed exploits that did not work were published, likely produced with LLMs, which the report says created significant verification work for defenders.

Google also sells security products, and the report pairs its findings with them: detection rules for one attack framework in Google Security Operations, a Mandiant hardening guide and the company's AI Threat Defense service for enterprise customers. The report gives no prices for them. Its findings are Google's own account, drawn from its telemetry and Mandiant's casework.

GTIG expects the tools to mature. It anticipates less human involvement and more stealth as the frameworks and the models behind them improve, and says it is highly likely that attackers will keep using them. Google's next quarterly update would show whether August's count of 22 zero-day exploits was a one-month spike or the start of a higher monthly rate.

Leave a Reply