An automated attack tool surfaced on an unprotected server, complete with Russian-language annotations, has been weaponizing artificial intelligence to compromise French media, entertainment and education organisations. Google's latest threat assessment from its Threat Intelligence Group and Mandiant division reveals a troubling ecosystem of AI-powered attack infrastructure, including stolen credentials for AI services and compromised Claude Code configurations.

According to Google's Monday disclosure, threat actors have deployed artificial intelligence to identify software vulnerabilities and launch attacks without any human intervention—a capability the company characterises as unprecedented in its threat monitoring. The discovery appears in the most recent quarterly AI threat report from the Google Threat Intelligence Group and Mandiant.

Researchers identified the system, designated exploit_agent, residing on an unprotected staging environment. The tool integrates with multiple cutting-edge and self-hosted AI models, enabling it to perform comparative analysis of software code across pre-patch and post-patch versions to identify probable security gaps. Once vulnerabilities are pinpointed, the system independently generates and deploys functional exploits.

During the previous quarterly assessment released two months earlier, the Threat Intelligence Group had reported no evidence of fully autonomous exploitation chains deployed against actual targets, according to reporting by SiliconANGLE.

The exposed infrastructure contained victim information alongside utilities for credential stuffing, deploying web-based backdoors and eliminating competing attackers' backdoors. Configuration documentation included annotations written in Russian, and victim profiles indicate targeting concentrated on French media, entertainment and educational institutions.

From patch to exploit

Google contends that patched vulnerabilities present a more tractable challenge for AI-driven exploitation than zero-day flaws. The reasoning is straightforward: attackers operate within a bounded problem space encompassing released code versions, security patches, technical documentation and published vulnerability research.

The analysis points to a pair of vulnerabilities affecting PaperCut print management systems that attackers combined in late August. Security researchers and the firm Huntress leveraged AI to construct functional attack code within twenty-four hours of public disclosure. Within four days, GreyNoise observed what appeared to be AI-orchestrated exploitation targeting hundreds of institutions.

Following the October patching of PHP vulnerabilities, hundreds of proof-of-concept code samples emerged that proved non-functional. Google's analysis suggests artificial intelligence generated a substantial portion of these ineffective exploits.

Attack frameworks in the wild

Multiple threat actors now incorporate publicly available AI-powered hacking frameworks into operational campaigns, according to Google's Threat Intelligence Group. These include profit-driven criminal syndicates and organisations suspected of conducting espionage on behalf of Chinese authorities. The CAIRN framework has been observed targeting entities within the American aerospace industry.

Google notes that these systems generate substantial forensic signatures and operational noise, rendering them comparatively straightforward to identify. The company anticipates that as these technologies mature, attackers will progressively minimise human involvement and refine their operational tradecraft to reduce detectability.

We believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally

Google Threat Intelligence Group

The threat landscape also encompasses attackers obtaining cloud infrastructure and AI service credentials to execute models using compromised accounts. In documented instances, fraudulent consumption charges reached $100,000 within a single twenty-four-hour window. One incident exploited a vulnerability in the LiteLLM gateway infrastructure. A group assessed as originating from North Korea modified Claude Code configuration files deployed within target organisations, ensuring that each activation would retrieve a malicious payload. Google reports having deactivated accounts and infrastructure associated with misuse of its Gemini platform.

Source: The Next Web