On November 16, 2026, Google Threat Intelligence Group (GTIG) and Mandiant reported finding an exposed staging server hosting exploit_agent, a tool that analyzed software patches, drafted exploits and executed them without human intervention. The finding describes automation of vulnerability analysis and exploitation—not a demonstrated attack carried autonomously from initial access through its final objective. GTIG’s tracker places the activity in patch-based exploitation.

What exploit_agent automated

Four steps connect comparing patched code with analyzing a vulnerability, drafting an exploit and executing it automatically.

GTIG said exploit_agent compared popular web-application code before and after patches, analyzed potential vulnerabilities, drafted exploits and ran them automatically. The framework used multiple frontier and local large language models.

The exposed server also contained data exfiltrated from targeted victims and scripts for internal reconnaissance, brute-force attempts, installing webshells and removing other attackers’ webshells. Server telemetry pointed to interest in French media, entertainment and education websites.

NeoTeo previously covered Google and Mandiant’s observations of agentic activity; this tracker describes the specific patch-analysis pipeline.

How much of an attack was automated?

The reported workflow focused on patched n-day vulnerabilities: flaws for which a patch has been released. By comparing code before and after a patch, a tool can analyze the change and develop an exploit from it. GTIG’s description concerns this patch-to-exploit process; it does not describe exploit_agent autonomously discovering zero-day vulnerabilities.

A vulnerability exploit can be automated without every stage of an intrusion being automated. GTIG separately described other agentic frameworks as handling discrete phases or tasks during intrusions. It said it had observed CAIRN and PentAGI in real-world activity, and experimentation with CyberStrikeAI and HexStrike AI. Those observations describe different tools and activities, not proof that each completed an entire intrusion autonomously.

Why patch timing and exploit verification matter

GTIG said agentic tools can compress the time between a patch disclosure and a weaponized exploit. It also said many AI-generated exploit claims are nonfunctional, creating extra validation work for defenders. Faster exploit development and checking whether a claimed exploit actually works are therefore separate pressures.

Patch quickly to narrow the exploit window, and validate exploit claims before treating them as working attacks.

The tracker reported broader vulnerability trends as context: vulnerability disclosures and exploits nearly doubled between January and August 2026. GTIG also reported an average of 8 zero-day exploits per month in 2025 and 11 per month in 2026, with 22 in August 2026. These are broad figures, not counts of exploit_agent activity.

GTIG said current agentic frameworks trade stealth for speed and can leave forensic traffic, recognizable identifiers and code-level traces. That gives defenders observable clues to monitor even as the tools automate more work.

A separate forecast for future attacks

In an article hosted by Yahoo Finance on November 18, 2026, 24/7 Wall St. attributed to GTIG a forecast of simple multi-agent attacks with little or no human interaction within one to three months, and sophisticated fully autonomous attacks within three to six months. These are projected timelines, not reports of attacks that had already occurred. Read the 24/7 Wall St. report.

Defensive measures described by GTIG

Google said SecOps had deployed behavioral and volumetric rule packs to detect ARTEX. Mandiant had also released a hardening guide for AI-enabled offensive tooling. These are specific detection and hardening efforts; GTIG’s report also emphasizes the continuing need to patch promptly and verify exploit claims.