Skip to content

Image: cyberscoop.com · rights & removal

Executive Summary

Russian-speaking hackers allegedly conducted what is believed to be the first end-to-end cyberattack fully automated by artificial intelligence, operating without human intervention at any stage of intrusion. This operation was documented in a quarterly report from Google’s Threat Intelligence Group and Mandiant incident response unit. Attribution is suggested by Russian-language comments found in the operation's code, including configuration instructions for automatic API key exploitation instead of manual input.
The attack targeted French websites across the media, entertainment, and education sectors on staging servers. The core of this automated system involved a mechanism dubbed “exploitagent,” which utilized multiple AI agents with both frontier and locally hosted language models to continuously analyze software code against vendor patches. This system automatically inferred changes, drafted exploits, and executed them. Investigators also discovered scripts used for internal reconnaissance, password guessing, and installing webshells that included the erasure of rival attackers' webshells.
This development signals a significant shift from current AI usage in cyber operations, where AI is primarily used to accelerate human tasks like writing code or summarizing data, to fully autonomous vulnerability discovery and exploitation. The report notes that threat actors are likely to continue using such frameworks globally due to their low barrier to entry and scaling potential.

Facts Only

* Russian-speaking hackers reportedly ran a cyberattack believed to be the first end-to-end AI attack without a human in the loop.
* The operation appeared in a quarterly report from Google’s Threat Intelligence Group and Mandiant.
* Attribution rests on Russian-language comments in the code, including instructions for "auto-exploitation."
* Victim data on staging servers pointed to targeting French websites in media, entertainment, and education.
* The centerpiece was a system named “exploitagent,” which chained multiple AI agents with language models.
* This system continuously compared code before and after vendor patches, inferred changes, drafted exploits, and executed them automatically.
* Investigators found scripts for internal reconnaissance, password guessing, and installing webshells.
* The tools were used to automate vulnerability discovery and exploitation without a human in the loop.
* Google observed open-source offensive frameworks like CAIRN and PentAGI in activity targeting U.S. aerospace organizations and Asian entities.
* Exploits appeared rapidly following vendor patches, with hundreds of purported exploits appearing within days.
* August saw 22 exploited zero-days, up from previous monthly averages.

Full Take

The narrative describes an inflection point where the mechanism for vulnerability research and intrusion moves from human-directed acceleration to autonomous execution. The significance lies not just in the use of AI for automation, but in creating a pipeline that is "unattended," which challenges existing security paradigms based on human decision-making loops. The system’s design—chaining agents to continuously test code updates and self-generate exploits—suggests a potential long-term threat where autonomous systems can operate at machine speed, rendering traditional human-centric defensive responses obsolete against highly adaptive threats.
The emphasis on speed, evidenced by the rapid appearance of nonfunctional or AI-generated exploits following patches, suggests that the primary risk is systemic latency in patching rather than the novelty of a single zero-day. The pattern highlights the tension between security innovation (like new AI models) and immediate threat response capabilities. Furthermore, the mention of established offensive frameworks being adapted by this automated approach suggests a convergence where publicly available tools become the foundation for autonomous action, which carries systemic risk regardless of attribution.
The implications for human agency revolve around whether the velocity of automated exploitation can be contained or countered before damage is done. If the standard response time is measured in hours or days due to patching cycles, and attacks are executed instantaneously by self-optimizing agents, the window for effective intervention shrinks dramatically. The focus shifts from mitigating individual vulnerabilities to securing the foundational integrity of the software supply chain and ensuring that defensive mechanisms can operate faster than autonomous offensive capabilities.
Bridge Questions:
If vulnerability discovery and exploitation become fully automated at machine speed, how must security architecture fundamentally change its dependency on human-in-the-loop verification? What mechanisms exist to establish meaningful accountability when autonomous agents execute actions derived from inferred code changes? How does the rapid evolution of autonomous offensive frameworks impact the viability of defensive patching strategies?

From the original · CyberScoop

Google said Monday that Russian-speaking hackers carried out what its analysts believe is the first cyberattack run end-to-end by artificial intelligence, without a human in the loop at any stage of the intrusion. The operation appears in a quarterly report from Google’s Threat Intelligence Group and its Mandiant incident response unit.
Read the full story at cyberscoop.com

Sentinel unavailable

The automated check of the source article's wording did not complete, so no result is shown.

This looks only at the wording of the original source article, not at this page's AI-written sections. A small local AI model made this estimate. It has not been checked against known human and machine texts, so treat it as provisional. It cannot show who wrote an article.

Google says Russian-speaking hackers ran the first fully automated AI cyberattack | Huntaegis