Google reports first fully autonomous AI-driven software exploits
Google's Threat Intelligence Group says attackers are now using large‑language‑model agents to discover and weaponise software bugs without human oversight.
Key takeaways
- Google identified an attacker‑controlled staging server running a multi‑agent framework called exploit_agent that autonomously scans patches and generates exploits.
- The AI‑driven pipeline reduced the window between a software fix and an exploit to minutes, as seen in an August 2026 PaperCut campaign.
- Google predicts simple autonomous attacks will become common within 1‑3 months and more sophisticated versions within 3‑6 months.
Google’s Threat Intelligence Group (GTIG) released its Q4 2026 AI Threat Tracker, documenting the first observed fully autonomous cyber‑attack pipeline that uses large language model (LLM) agents to locate, weaponise and deploy software vulnerabilities without any human intervention.
Autonomous vulnerability research pipeline
According to the GTIT report, investigators uncovered an exposed staging server that hosted a multi‑agent framework dubbed **exploit_agent**. The system continuously performed differential analysis of code bases before and after patch releases, identified changes that introduced new flaws, and automatically generated working exploit code. The framework connects to multiple LLMs, allowing it to read release notes, version diffs and public proof‑of‑concepts, then write and execute attacks on internet‑facing servers.
Real‑world impact and speed
The report cites an August 2026 campaign against PaperCut NG/MF software where the AI pipeline achieved domain‑administrator control within minutes of the patch’s public release, according to GreyNoise. Google notes that vulnerability disclosures more than doubled between January and August 2026, while exploited zero‑day counts rose from an average of eight per month in 2025 to eleven per month in early 2026, spiking to 22 in August.
Broader implications and response
Google warns that the speed of AI‑driven attacks shortens the remediation window for organizations, especially those slow to apply updates. The report also highlights a trend called “LLMjacking,” where threat actors steal AI credentials and cloud resources to resell them on underground markets such as APIMart and JetNight. To counter the threat, Google is promoting its AI Threat Defense service, which leverages Gemini models to proactively discover and patch vulnerabilities before malicious agents can exploit them.
What to watch next
GTIG expects simple autonomous attacks to appear widely within the next one to three months, with more advanced multi‑stage operations emerging in three to six months. Observers should monitor the adoption of open‑source frameworks like ZeroHollow, Agent Zero and the Chinese‑language phishing platform Jieyou, as well as any uptick in AI‑credential theft incidents that could fuel further automated campaigns.
FAQ
What is the exploit_agent framework?
Exploit_agent is a multi‑agent system found on an attacker’s staging server that uses large language models to compare software versions, identify flaws and generate exploit code without human oversight.


Discussion 0