Google Thwarts AI-Developed Zero-Day Exploit

For the first time, Google has identified and neutralized a zero-day exploit that appears to have been crafted with the help of AI. A report from the Google Threat Intelligence Group (GTIG) reveals that “prominent cyber crime threat actors” were gearing up to leverage this vulnerability for a large-scale exploitation event, which would have enabled them to circumvent two-factor authentication on an unspecified “open-source, web-based system administration tool.”

Researchers at Google discovered clues within the Python script associated with the exploit that suggested AI involvement, such as a “hallucinated CVSS score” and a “structured, textbook” format typical of large language model training data. The exploit capitalizes on a “high-level semantic logic flaw where the developer hardcoded a trust assumption” into the platform’s 2FA system. This development comes on the heels of ongoing discussions about the capabilities of AI models focused on cybersecurity, like Anthropic’s Mythos, and a recently revealed Linux vulnerability that was uncovered with AI assistance.

This marks the first instance where Google has found evidence of AI being implicated in such an attack, although researchers clarify that they “do not believe Gemini was used.” Google successfully “disrupted” this specific exploit but warns that hackers are increasingly employing AI to discover and exploit security weaknesses. The report also highlights AI as a target for attackers, noting that “GTIG has observed adversaries increasingly target the integrated components that grant AI systems their utility, such as autonomous skills and third-party data connectors.”

Additionally, the report outlines how hackers are utilizing “persona-driven jailbreaking” techniques to prompt AI into identifying security vulnerabilities, including an example where the AI is instructed to act as a security expert. Hackers are also supplying AI models with entire repositories of vulnerability data and employing OpenClaw in ways that indicate a desire to refine AI-generated payloads in controlled environments to enhance exploit reliability before deployment.