Key Highlights
- Google uncovered an autonomous hacking tool capable of identifying software flaws and launching exploits without human intervention.
- Discovered on an exposed staging server, the tool—named “exploit_agent” by Google—connects to commercial and “local” large language models.
- The automated agent analyzes security patches by diffing code before and after fixes to automatically write and execute functional exploits.
Discovery of the Autonomous “exploit_agent”
Google announced on Monday that its security researchers have uncovered an automated hacking tool designed to discover software vulnerabilities and exploit them entirely on its own. The discovery, documented in the latest third-quarter report from the Google Threat Intelligence Group, represents a tangible example of the emerging cybersecurity risks posed by autonomous artificial intelligence agents operating without direct human oversight.
Security analysts identified the mechanism on an insecure staging server—an infrastructure system routinely utilized by cyber threat actors to store, configure, and stage attack software—that had been inadvertently left accessible to the public internet. Google designated the automated system “exploit_agent”.
How the AI-Driven Exploit Engine Operates
The architecture of “exploit_agent” relies on integrations with multiple large language models (LLMs), the core technology powering modern generative AI platforms and chatbots. According to Google, the tool connected to a combination of leading commercial foundation models as well as models categorized as “local”.
Automated Patch Diffing and Exploit Generation
Once deployed, the agent continuously monitored widely used web software for security updates and vendor-issued patches. Upon the release of an update, the tool automatically compared the software’s codebase before and after the fix was applied—a technique known in security analysis as patch diffing. By evaluating precisely which lines of code were modified by developers to resolve a vulnerability, the system isolated the original flaw. Immediately after mapping the security weakness, “exploit_agent” generated a tailored exploit payload and executed the attack automatically, removing the human lag typically required to reverse-engineer fixes.
Why This Matters
The discovery of “exploit_agent” signals a shift in the timeline between vulnerability disclosure and weaponization. Traditionally, when software vendors release patches, security administrators have a brief window to apply updates before threat actors can manually dissect the fixes and develop working exploit code. By leveraging LLMs to analyze code diffs, synthesize functional exploits, and launch attacks autonomously, AI-driven tooling drastically reduces that defensive window, highlighting how modern machine learning models can be repurposed to streamline offensive cyber operations.
Frequently Asked Questions
What is “exploit_agent”?
“exploit_agent” is a specialized hacking tool identified by Google Threat Intelligence Group that connects to commercial and “local” large language models to autonomously discover software vulnerabilities and execute exploits without human direction.
How does the tool find security flaws in software?
The tool tracks updates released for popular web software and compares the code before and after a patch is issued. By identifying the exact code changes made to address a problem, the AI pinpoints the underlying vulnerability and automatically drafts code to exploit it.
Where was the AI hacking tool discovered?
Google researchers discovered the tool residing on an exposed staging server—a backend computer utilized by threat actors to test and house their tools—that had been left open to the internet.




