Google says its Big Sleep AI agent recently discovered a critical SQLite vulnerability and thwarted threat actors’ efforts to exploit it in the wild.
Big Sleep, an AI agent developed by Google’s DeepMind and Project Zero teams, is designed to actively search for unknown vulnerabilities in software.
Google claimed in November 2024 that Big Sleep had managed to find its first real-world vulnerability, an exploitable buffer overflow in the widely used open source database engine SQLite.
The tech giant said at the time that its researchers had attempted to find the same vulnerability using fuzzing, but they failed to accomplish the task.
In the case of that SQLite vulnerability, it was discovered in a version of the software that had yet to be released, meaning that users were not at risk.
However, in a blog post published on Tuesday, Google said Big Sleep recently discovered another SQLite vulnerability that was “only known to threat actors and was at risk of being exploited”.
The vulnerability, tracked as CVE-2025-6965, has been described as an issue related to the fact that the number of aggregate terms could exceed the number of available columns, leading to memory corruption. The vulnerability was patched in late June with the release of version 3.50.2.
No other details are available, but memory corruption vulnerabilities can typically lead to arbitrary code execution, privilege escalation, data leakage, or denial of service.
“Through the combination of threat intelligence and Big Sleep, Google was able to actually predict that a vulnerability was imminently going to be used and we were able to cut it off beforehand,” Google said. “We believe this is the first time an AI agent has been used to directly foil efforts to exploit a vulnerability in the wild.”
Learn more about AI-powered security solutions at
SecurityWeek’s 2025 AI Risk Summit
SecurityWeek has asked Google to share additional technical details, but the company has refused to do so.
It’s unclear what information had been given to Big Sleep by threat intelligence experts, and how the company determined that the vulnerability was at risk of being exploited.
Potentially critical SQLite vulnerabilities have come to light every once in a while, but there do not appear to be any reports describing the in-the-wild exploitation of such flaws.
For instance, CISA’s Known Exploited Vulnerabilities (KEV) catalog does not include any SQLite flaws, although the government agency’s list is known to be incomplete.
Google also announced on Tuesday that it’s donating data from its Secure AI Framework to the Coalition for Secure AI (CoSAI), an initiative aimed at tackling the cybersecurity risks associated with AI. This will “help accelerate CoSAI’s agentic AI, cyber defense and software supply chain security workstreams”, the company said.
Related: Grok-4 Falls to a Jailbreak Two Days After Its Release
Related: Google Gemini Tricked Into Showing Phishing Message Hidden in Email
