Google: exploit_agent Automated Hacking With No Human in the Loop
· Cortex Solutions

Google's threat intelligence team says attackers used exploit_agent, an AI framework that turns software patches into working exploits with no human in the loop.
Google Threat Intelligence Group (GTIG), working with Mandiant, says it has identified threat actors using large language model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time. The finding appears in the quarterly GTIG AI Threat Tracker published on the Google Cloud blog, dated November 17, 2026, and covered by SiliconANGLE, CyberScoop and BleepingComputer on November 16, 2026.
What is exploit_agent?
According to Google's report, researchers found an exposed staging server belonging to the attackers, hosting a multi-agent vulnerability research and weaponization framework dubbed exploit_agent. The report describes how it works:
- It integrates with several frontier and local LLMs; the report does not name any specific model.
- It runs continuous differential analysis of popular web application software, comparing code before and after a security patch is announced.
- It analyzes the differences for potential vulnerabilities, then drafts the exploit code.
- It executes the exploit automatically, without human intervention.
Alongside the framework, the same server held data exfiltrated from victims, plus scripts for internal reconnaissance, password brute-forcing and installing webshells, and even scripts for removing webshells planted by other attackers. Victim telemetry, according to Google, points to targeting of French websites tied to the media and entertainment industries, as well as the education sector.
The tool's configuration contained Russian-language comments referring to LLM API keys for auto-exploitation. CyberScoop noted, however, that researchers did not name a specific group or state behind the activity.
Why is this an important shift?
The news matters because the gap between publishing a security patch and exploiting the flaw it fixes is narrowing. The framework does not hunt for unknown flaws; it reads published patches to infer what was fixed, then attacks sites that have not yet been updated.
The report gives figures that show how fast the landscape is moving: exploited zero-days averaged 8 a month in 2025 and 11 a month in 2026, then jumped to 22 in August 2026 alone. CyberScoop also reported that after PHP patches were released on October 21, 2026, hundreds of purported exploits appeared, most of them nonfunctional and likely AI-generated, which distracts defenders.
Other tools in the picture
exploit_agent was not the only tool in the report. According to Google, the open-source tools CAIRN and PentAGI were observed deployed in the wild: CAIRN was used against US aerospace targets, and a customized platform built on PentAGI was used from late August to early October 2026 for reconnaissance and scanning of organizations in Asia. Google also saw experimentation with CyberStrikeAI and HexStrike AI.
The report assesses that these frameworks currently trade stealth for velocity, and says it is highly likely that threat actors will continue to use them in campaigns targeting organizations globally. SiliconANGLE reported that Google disabled the accounts and assets tied to misuse of its services.
What does this mean for Saudi businesses?
Saudi businesses that run websites, online stores and customer portals are not out of reach of this kind of attack, because it targets popular web software wherever it is deployed. The following are practical recommendations from the Cortex Solutions team, not from Google's report:
- Treat security updates as urgent: if an attacker can turn a patch into an attack automatically, delaying updates for weeks leaves the door open.
- Inventory everything you have online: websites, plugins, control panels and forgotten staging servers, because an asset you do not know about does not get updated.
- Keep plugins and third-party components on your sites to what you actually need, and track their versions.
- Enable monitoring and alerts on your servers to catch unusual files or password-guessing attempts early.
- Protect the AI API keys your business uses, and never leave them in exposed files.
- Agree with your hosting provider or technical team on a fixed timeframe for applying critical updates.
The bottom line: AI is no longer just a productivity tool; it is also speeding up attacks. Businesses that update their systems quickly and know their digital assets well will be in a much stronger position than others.
Sources: the GTIG AI Threat Tracker on the Google Cloud blog (November 17, 2026), and coverage by SiliconANGLE, CyberScoop and BleepingComputer (November 16, 2026).