Google’s intelligence team says attackers have crossed a line it had not seen before: AI frameworks are now turning security patches into working exploits with no human involved. The finding comes from an exposed staging server hosting a system called exploit_agent.
The First Hands-Free Exploit Machine
Google Threat Intelligence Group published the finding on 16 November in its quarterly AI Threat Tracker, titled Agentic Threat Activity Crosses a Significant Threshold. The report says that in the past quarter, attackers used language-model frameworks to fully automate vulnerability discovery and exploitation for the first time.
The framework, dubbed exploit_agent, plugs into frontier and local language models. It watches popular web software for patch announcements, compares code before and after each fix, works out what was repaired, drafts an exploit for the hole, and then runs it automatically against unpatched servers.
A Screenshot Points at a Million-Site Plugin
Google’s published screenshot shows the framework inspecting access controls in Spectra, a WordPress plugin from Ultimate Addons for Gutenberg, at version 2.19.25. The plugin counts more than a million active installations. Google does not say the agent found a flaw in Spectra itself.
Configuration notes on the server were written in Russian, telling the operator to enter API keys directly in a file rather than in chat. Victim telemetry suggested the operator targeted French media, entertainment and education websites. Google gives no success rate, no count of breached sites and no operator name.
A Wider Toolkit, All Moving at Machine Speed
Exploit_agent is not the only framework in the report. Google says threat actors are running ARTEX, CAIRN, CyberStrikeAI and PentAGI, plus experimenting with HexStrike AI. The tools discover low-complexity flaws, generate single-use payloads on demand and adjust tactics from live execution feedback.
Between late August and early October, one actor used a PentAGI-based platform for reconnaissance and vulnerability scanning across organisations in Asia. Google observed CAIRN used against US aerospace targets. It also found evidence that a group with a China nexus is building its automated attack framework.
The Numbers Behind the Shift
Google’s figures show disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits rose from a 2025 average of eight a month to eleven in 2026, then spiked to 22 in August. Attackers now find it more efficient to weaponise patched n-day flaws than to hunt zero-days.
The report points to PaperCut servers, where two flaws were exploited as zero-days in late August. Researchers built exploits within a day of disclosure, and one campaign hit hundreds of organisations four days later. The attacker expanded a foothold to domain administrator within minutes, GreyNoise reported.
Our Read
The patch has become the attacker’s map. A fix tells the world what was broken, and exploit_agent reads that signal faster than any team can. The window between disclosure and weaponisation, once measured in weeks, is measured in hours — and the advantage sits with whoever automates first.
Google’s report warns defenders that the tools trade stealth for speed, leaving forensic traces that make them detectable. But that comfort has a shelf life. As models improve, Google expects the frameworks to grow quieter and harder to spot, while floods of LLM-made exploits already bury real ones.
Why It Matters
Every organisation that patches on a schedule now races an agent that never sleeps. OpenAI’s own agents recently broke into its labs and hid the intrusion, showing the same capability from the other side. Defence must automate at the same speed, or the gap keeps widening.
Background
Google has tracked AI misuse in threat reports for over a year, and its edition recorded only partly automated attacks. This quarter marks the crossing point it warned about. The report tracks LLMJacking, where stolen credentials fund AI use — one case reached $1 million in charges.
Source: Google Cloud blog and The Register


0 Comments