Google Threat Intelligence Group (GTIG) researchers report that a zero-day exploit aimed at a popular open-source web administration tool was probably crafted using AI.

This exploit allowed attackers to sidestep two-factor authentication (2FA) safeguards in an unnamed, widely adopted open-source web-based administration platform.

Although authorities halted the attack prior to any widespread deployment, the event underscores how cybercriminals are increasingly depending on AI for detecting vulnerabilities and building exploits.

Google expresses strong confidence that the attacker employed an AI model to uncover and operationalize the flaw, based on the Python exploit code's structure and details.

"The script is filled with instructional docstrings, features a fabricated CVSS score, and follows a rigid, textbook-like Python style typical of large language model (LLM) training datasets," GTIG explains in a new report.

While the exact LLM behind this malicious activity is unknown, Google confirms that its Gemini model played no role.

Supporting LLM involvement is the vulnerability's character: a complex semantic logic error that AI tools are adept at spotting, unlike traditional memory corruption or sanitization flaws detected via fuzzing or static checks.

Google informed the software vendor of the grave risk and collaborated on swift measures to neutralize the threat.

"GTIG has for the first time detected a threat actor wielding a zero-day exploit we assess as AI-developed," the researchers declare.

In addition to this incident, Google points out that groups from China and North Korea - including APT27, APT45, UNC2814, UNC5673, and UNC6201 - continue leveraging AI for vulnerability research and exploit creation, aligning with patterns from a February analysis.

Russia-affiliated operatives have similarly deployed AI-generated decoy scripts to conceal malware like CANFAIL and LONGSTREAM.

Google draws attention to the Russian "Overload" campaign, in which social engineers applied AI voice synthesis to pose as legitimate journalists in fabricated videos pushing anti-Ukraine propaganda.

The report also spotlights the Android PromptSpy backdoor, previously detailed by ESET this year, which incorporates Gemini APIs to enable hands-free device manipulation.

Researchers identified a self-operating component called "GeminiAutomationAgent," relying on a preset prompt to facilitate malware-driven automated interactions with the device.

This prompt establishes a non-threatening identity to evade the LLM's protective mechanisms, ultimately computing user interface boundaries for diverse interaction methods.

Moreover, PromptSpy harnesses AI functions to mimic authentication inputs such as lock patterns or PINs on the device, according to Google.

Google cautions that adversaries are now systematically accessing advanced AI services through automated account generation, proxy networks, and shared credential systems.