Google Threat Intelligence Group (GTIG) found an exposed staging server running exploit_agent, a multi-agent framework that compares web application code before and after security patches, drafts exploits, and runs them without human intervention (GTIG). GTIG calls it the first fully automated discovery-to-exploitation pipeline it has seen in the wild. The server also held data stolen from victims, and its records point to French media, entertainment, and education websites.
What happened
The framework integrates with several frontier and local models and runs continuous differential analysis of popular web application software. One captured analysis inspected the access check on a REST route in the Spectra plugin for WordPress, version 2.19.25. GTIG does not say which flaws were exploited against which victims.
A configuration file holds LLM API keys under Russian-language comments that call the setup an “exploit_agent pilot,” which is why GTIG suspects a Russian-speaking operator (The Register). Alongside the stolen data, the server held scripts for internal reconnaissance, brute-forcing, installing web shells, and removing other attackers’ web shells.
GTIG places this in a wider shift: it counted 22 zero-days exploited in August 2026, up from an average of 8 a month in 2025. It also notes these frameworks currently trade stealth for speed and leave forensically noisy traces.
How it could have been caught
When an agent can turn a patch into an exploit in one run, the gap between a vendor’s fix and an attack shrinks to roughly the time it takes to apply that fix. A site still running the old plugin has to catch the exploit by what it does. A web server’s PHP workers answer requests and query their database, and the first behavior a baseline would have flagged is a worker spawning a shell.
What this looks like in Trace
Here is one exploitation traced on a cloud-hosted WordPress site of the kind GTIG describes. The plugin analysis, web shells, reconnaissance scripts, and data theft come from GTIG. The site, the exploited flaw, ancestry, hostnames, ports, and timings are illustrative.
Network trace. Workload web/wordpress, the site’s PHP application tier.
- T+0:00nginx → php-fpm (WordPress REST request)
- Destination
- wp-db:3306
- Baseline
- Known
- T+0:03php-fpm → sh
- Baseline
- Anomaly, new ancestry
- T+0:20sh → mysqldump
- Destination
- wp-db:3306
- Baseline
- Anomaly, bulk export
- T+0:41sh → curl
- Destination
- staging server:443
- Baseline
- Anomaly, never-seen destination
The operator’s models answered on its own infrastructure, so no AI proxy at the victim saw them. The speed GTIG describes works against the attacker here, because the shell, the export, and the upload land within a minute and form one case.
Traceback. At T+0:03 the triage agent opens a case and follows the chain:
upload to a new host → mysqldump → sh → php-fpm → REST request to a plugin route, with no deploy, plugin update, or admin session behind the shell
The chain ends at an anonymous request to a plugin, and that dead end is the finding.
Verdict: an autonomous agent exploiting an unpatched plugin. The case reaches the site’s operators within seconds of the shell starting, with the route, the shell, and the destination named, before the export leaves the network.
This analysis reconstructs the incident from public reporting and walks through where Trace would have raised the first alert inside the cloud hosting a targeted website. Detection times are illustrative.