
Google Says AI Now Builds and Runs Exploits With No Human Involved
Google says attackers ran a fully autonomous AI exploit framework, exploit_agent, against real targets as zero-day counts hit 22 in August, up from 11 a month.
- By
- Nathan Brooks
- Filed
- Channel
- AI & Compute
- Read
- 3 min read
Zero-day exploits jumped from an average of 11 a month in 2026 to 22 in August, and Google now says attackers used an AI framework to find and exploit software flaws entirely without human direction. Google's Threat Intelligence Group (GTIG) reported on Monday that it "identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time." The team calls the shift "a paradigm shift towards intrusions occurring at machine speed." Just two months ago, GTIG's previous report had found no fully autonomous attack pipelines in the wild. That gap has closed.
What does exploit_agent actually do?
GTIG discovered the framework, named exploit_agent by its operator, on an exposed staging server. It connects to "multiple frontier and local LLMs" and runs continuous differential analysis of popular web application software. The tool compares code before and after a patch ships, works out which flaw the patch fixed, drafts an exploit, and executes it "without human intervention." One worked example in the report shows the tool inspecting an access-control check in a version of the Spectra plugin, also known as Ultimate Addons for Gutenberg, for WordPress.
The staging server held more than the framework. It also contained data exfiltrated from victims plus scripts for reconnaissance, password brute-forcing, webshell installation, and the removal of webshells left by rival attackers. Victim telemetry pointed at French websites in media, entertainment, and education. GTIG declined to attribute the operator to a country. The configuration file carries comments in Russian, including one translated as "LLM API keys for auto-exploitation (exploit_agent pilot)."
Why patches now help attackers
GTIG argues that AI makes n-days — flaws that already have patches — a more efficient target than zero-days, because attackers "can focus research on a finite set of source materials," whereas zero-day hunting "requires casting a wider net." The practical consequence: AI is "shrinking the patch window for organizations."
The PaperCut print-server flaws exploited as zero-days in late August, CVE-2026-81578 and CVE-2026-82078, illustrate the speed. Researchers and Huntress produced working exploits with AI within a day of disclosure. GreyNoise reported a campaign four days after disclosure, likely built on AI-generated exploits, in which an attacker moved from an initial foothold to domain administrator within minutes. GTIG also cited the PHP flaws behind the PHP2Shell attacks, where early patched code circulated publicly before the official fix on October 21.
Which other AI-assisted intrusions did Google find?
The report catalogues several parallel cases:
- Attackers deployed CAIRN, an open-source multi-agent penetration-testing framework, in intrusions that included US aerospace targets.
- One actor ran a customized platform built on the open-source PentAGI framework to scan organizations in Asia from late August to early October.
- A cluster tracked as MIDNIGHT NEPTUNE approached cryptocurrency companies in North America and Europe through fake recruiter profiles, then modified .claude/settings.json files in internal code repositories so that launching Anthropic's Claude Code downloaded a backdoor named SOMBERMEME.
Google also policed its own infrastructure. GTIG disabled accounts and assets tied to misuse of the Gemini API, including a commercial contractor's platform built to manage more than 100,000 social media accounts. Google DeepMind used the findings to strengthen its classifiers.
Does autonomy cut both ways?
For now, yes. GTIG notes that unrestricted autonomous tools leave a noisy trail, which makes them easier to detect. Defenders are arming themselves with the same models. Anthropic and Google researchers found the critical HTTP/2 flaw that Apache patched last week, and OpenAI launched AI agents that draft security patches on November 13.
The equilibrium is temporary. If exploit quality improves faster than detection, the patch window that AI is already shrinking will close faster than most organizations can patch.
Source: mlq.ai
More from Nathan Brooks
Show full bio
Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.
1,010 articles
Related articles
attackers-run-open-source-ai-pentest-agent-antifox-to-steal-aws-secrets
Attackers Run Open-Source AI Pentest Agent AntiFox to Steal AWS Secrets
openai-s-gpt-6-cyber-reproduced-1-724-kernel-bugs-as-agent-tools-launch
OpenAI's GPT-6 Cyber Reproduced 1,724 Kernel Bugs as Agent Tools Launch
hackers-ran-espionage-campaign-with-ai-agent-built-on-deepseek
Hackers Ran Espionage Campaign With AI Agent Built on DeepSeek


