Google Blocks AI-Developed Zero-Day Exploit

What Google says it found

Google’s researchers said the exploit was written in Python and included details that suggested artificial intelligence had helped create it. Those clues included a “hallucinated CVSS score” and formatting that appeared “structured” and “textbook” in a way consistent with large language model training data.

The flaw itself was not described as a simple coding typo. Google said it involved “a high-level semantic logic flaw where the developer hardcoded a trust assumption” inside the platform’s two-factor authentication system. In plainer terms: the system trusted something it should not have trusted, and attackers found a way to make that matter.

Google said it was able to “disrupt” the exploit before the planned large-scale attack unfolded. The company did not name the affected administration tool.

Why the AI angle matters

The report marks the first time Google says it has found evidence that AI was involved in a zero-day attack of this kind. The researchers added that they “do not believe Gemini was used,” which is useful clarification, if also a reminder that the AI ecosystem is now large enough for brand attribution to become part of incident response.

The finding arrives during a period of growing concern about cybersecurity-focused AI systems. Security researchers and technology companies have been debating the capabilities of models designed or adapted for offensive security work, including Anthropic’s Mythos. A recently disclosed Linux vulnerability discovered with AI assistance has added more fuel to that discussion.

Google’s broader point is not that AI has suddenly created cyber crime. That already had a full calendar. The concern is that attackers are increasingly using AI to identify weaknesses, generate exploit code, and improve their chances before launching attacks.

How attackers are using AI tools

The Google Threat Intelligence Group said it has observed hackers using “persona-driven jailbreaking” to push AI systems into helping with vulnerability research. One example involved prompting an AI tool to pretend it was a security expert, a familiar tactic in attempts to bypass model safety limits.

The report also said attackers are feeding AI models large repositories of vulnerability data. That can help systems identify patterns, suggest exploit paths, or generate code more quickly than a human working from scratch.

Google also pointed to the use of OpenClaw in ways that suggested “an interest in refining AI-generated payloads within controlled settings to increase exploit reliability prior to deployment.” In other words, attackers are not only asking AI for ideas. They are testing and polishing the output before using it.

AI systems are becoming targets too

Google warned that AI is not only a tool for attackers. It is also becoming a target. The report said the group has seen adversaries increasingly focus on the integrated components that make AI systems useful, including autonomous skills and third-party data connectors.

That matters because modern AI systems often do more than answer questions. They may connect to external services, pull in private data, or take actions across software environments. Each of those connections can become another place for attackers to look for leverage.

For security teams, the message is fairly direct: AI-assisted attacks are moving from theory to operational reality. Google says it stopped this one. The next one may not arrive with such convenient warning signs.