MediumVulnerabilityMonitoring

Google GTIG: AI agent exploit_agent turns patches into exploits

Google GTIG reports an AI agent that turns security patches into working exploits unaided; apply web application patches promptly.

Published
Updated
Exploitation
Yes, per Google GTIG, an exposed staging server held stolen victim data, and the CAIRN framework was used in a US aerospace intrusion

Live scores

CVE-2026-81578
EPSS
0.847
very likely to be attacked soon
VEDAS
0.978
strong exploit evidence
On CISA KEV
Yes, since 2026-08-31
Detection rule
Nuclei · Suricata
EPSS (chance of attack)
100.1Added to CISA KEV 2026-08-31CVE-2026-81578 EPSS over time2026-082026-11-17
VEDAS (exploit evidence)
100.5Added to CISA KEV 2026-08-31CVE-2026-81578 VEDAS over time2026-082026-11-17

Thin horizontal line: the “likely exploited” threshold (EPSS 0.10, VEDAS 0.5). Shaded “no data” areas are periods when cve-scores did not publish. The vertical line is the day CISA added it to its exploited list.

VEDAS flagged this bug about 2 days before CISA confirmed attacks.

Live scores

CVE-2026-82078
EPSS
0.630
very likely to be attacked soon
VEDAS
0.975
strong exploit evidence
On CISA KEV
Yes, since 2026-08-31
Detection rule
No public rule yet
EPSS (chance of attack)
100.1Added to CISA KEV 2026-08-31CVE-2026-82078 EPSS over time2026-082026-11-17
VEDAS (exploit evidence)
100.5Added to CISA KEV 2026-08-31CVE-2026-82078 VEDAS over time2026-082026-11-17

Thin horizontal line: the “likely exploited” threshold (EPSS 0.10, VEDAS 0.5). Shaded “no data” areas are periods when cve-scores did not publish. The vertical line is the day CISA added it to its exploited list.

VEDAS flagged this bug about 2 days before CISA confirmed attacks.

What happened

On 16 November 2026 Google published its fourth-quarter 2026 AI Threat Tracker. The Register reports that GTIG found an exposed staging server hosting a multi-agent vulnerability research and weaponization framework called exploit_agent. It was configured for continuous differential analysis of “popular web application software”: comparing code before and after security updates, identifying likely vulnerabilities, developing exploits and executing them automatically. The Register

Details reported by the outlets:

  • Claimed first. GTIG describes this as its first observation of attackers fully automating vulnerability research and exploitation with no human in the loop. Hackread notes this is a finding from Google’s investigations, not a claim that nobody has tried it before. Hackread
  • The server. It also held data stolen from victim organizations and several post-compromise scripts, including internal reconnaissance, brute-forcing, webshell installation and removal of other attackers’ webshells. The Register
  • Targets. Victim telemetry suggested interest in French websites in media, entertainment and education. The Register
  • Attribution is soft. The Register describes a suspected Russian-speaking operator. Cyber Security News says Russian comments in the setup files pointed to a pilot project. Hackread cautions that language alone does not establish the operator’s nationality. Cyber Security News, Hackread
  • Limits. Hackread says the report gives no success rate and does not establish that every generated exploit worked. Hackread

The report also covers open-source agentic frameworks used for multi-stage intrusions:

  • Frameworks named. ARTEX, CAIRN, CyberStrikeAI and PentAGI are named by The Hacker News. Cyber Security News adds HexStrike AI and Agent Zero. The Hacker News, Cyber Security News
  • CAIRN in the aerospace case. Nextgov reports CAIRN was used in an intrusion into the US aerospace sector. Google did not name the organization, say what was taken, or attribute the hack to a government. It is not clear when the intrusion happened. Nextgov
  • PentAGI in Asia. Cyber Security News says a customized PentAGI version scanned organizations across Asia between late August and early October 2026. It says CyberStrikeAI and HexStrike AI were only observed in testing. Cyber Security News
  • Actors. Google says financially motivated actors and suspected China-nexus espionage actors have experimented with these tools. It judges it “highly likely” they will keep using them. The Register
  • Trade-off. Google says these tools trade stealth for speed and leave noisy traffic and code traces. Cyber Security News

PaperCut and patch diffing. The Hacker News says LLM-driven patch analysis was observed in the case of PaperCut NG/MF flaws “earlier this September”. Cyber Security News names two PaperCut flaws, CVE-2026-81578 and CVE-2026-82078, exploited as zero-days in August 2026. It cites a GreyNoise report of an AI-assisted campaign that reached domain admin within minutes. The two outlets differ on timing, and neither says exploit_agent was involved. Treat the link to exploit_agent as not confirmed. The Hacker News, Cyber Security News

The report also describes LLM credential theft resold through proxy services, North Korea-linked fake recruiter lures, and AI-assisted influence operations. Those are outside the scope of this advisory.

Am I affected?

There is no single product or version to check. The sources do not say which web application software exploit_agent targeted, so that is not confirmed.

  • Organizations running internet-facing web applications have the most exposure to the patch-to-exploit window. Hackread
  • The reported victim interest was French media, entertainment and education sites, a US aerospace organization (CAIRN), and organizations across Asia (PentAGI). The Register, Nextgov, Cyber Security News
  • Compromise indicators, such as hashes and IP addresses, are not included in the sources reviewed. The Google report itself was not among them.

What to do now

  1. Work out which applications need updates and apply security fixes promptly, especially on systems reachable from the internet. Hackread
  2. If an update has to wait, limit access to the affected application. Hackread
  3. Check for signs of an earlier compromise, including webshells on web servers, since patching does not evict an attacker who is already in. Hackread
  4. Follow the defensive guidance Cyber Security News summarizes from Mandiant: lock down admin systems, use short-lived credentials, limit access, split networks, and watch for sudden spikes in API and identity activity. Cyber Security News
  5. If you use Google SecOps, note that Google has released behavior-based rules to detect ARTEX. Cyber Security News
  6. For AI services, set budget alerts, scan code for leaked keys, and treat unexpected AI service activation as a serious warning. Cyber Security News

Sources