Google Threat Intelligence Group examined an exposed staging server running an AI-assisted framework that compares web-application code before and after patch announcements, writes exploits and runs them automatically. For teams responsible for internet-facing web apps, the release of a security fix can also give this sort of operator a map of what to attack on servers still awaiting the fix. GTIG calls the framework exploit_agent and says it executes drafted exploits “without human intervention.”

The server held data exfiltrated from targeted victims and scripts capable of internal reconnaissance, brute-forcing, installing webshells and removing other attackers’ webshells, according to GTIG. These are available post-compromise capabilities on the staging server; check your own process and file records to determine whether any ran on a particular host. Victim telemetry led GTIG to identify French media and entertainment websites and education sites as particular interests of the operator. The report was published November 16, 2026.

Patches Feed the Agent

A patch diff is a shopping list for anyone searching for a newly fixed weakness. GTIG says exploit_agent continuously compares versions of popular web-application software around patch announcements. The framework uses frontier and local language models to inspect the differences, identify potential flaws, draft exploits and execute them. That sequence takes human approval out of the exploitation loop. Human attackers still choose targets and objectives.

The report shows the framework inspecting an access-control check in Spectra (Ultimate Addons for Gutenberg), a WordPress plugin, at version 2.19.25. Its excerpt names a local file, v21925.php, and the route /wp/v2/spectra-popup. GTIG observed analysis of this check. Teams with that plugin should compare their installed version with the maintainer’s security notices and retain the version and route when examining relevant access logs; correlate requests to that route with local file changes to investigate what happened on their own sites.

GTIG also describes autonomous penetration-testing frameworks in live intrusions. It observed CAIRN and PentAGI deployed in the wild, including CAIRN activity against targets in the US aerospace sector; it describes experimentation with CyberStrikeAI and HexStrike AI separately. GTIG says unrestricted autonomous runs trade stealth for speed, leaving forensic traffic, recognizable identifiers and code-level traces. That visibility gives defenders somewhere to look when an exposed service receives a burst of unfamiliar requests soon after a patch ships.

PaperCut Shows the Patch Window

The Google report cites a separate PaperCut episode from late August: a chain involving an authentication bypass, CVE-2026-81578, and unsafe dynamic class loading, CVE-2026-82078. The PaperCut-assigned record for the bypass describes remote unauthenticated requests reaching administrative backend actions before access validation completes. Once configuration parameters are changed, the second record says the database connector can load attacker-selected Java classes from its classpath under the PaperCut server process. Together the flaws allow remote code execution, according to GTIG.

GTIG cites GreyNoise’s report of a campaign against hundreds of organizations four days after disclosure; GreyNoise says the attackers likely used LLMs in exploit development and targeting. GTIG cites PaperCut as a separate exploitation case; it identified exploit_agent on a staging server. The CVE records carry CISA’s known-exploited entries for both vulnerabilities. PaperCut MF and NG releases before 24.1.10 in the 24.x line, before 25.0.13 in the 25.x line, and before 26.0.5 in the 26.x line are marked affected in both records. The two records describe the first fixed builds for those lines as 24.1.10, 25.0.13 and 26.0.5, respectively.

The PaperCut attacks arrived four days after disclosure, putting pressure on teams still running affected builds. The automation GTIG saw on the staging server targets the same kind of gap: a defender has a newly fixed code base to deploy while an attacker can compare it against the older version still online. The report also says many claimed LLM-generated exploits fail. Prioritize observed traffic and installed versions over an untested proof-of-concept claim.

Check the Exposed Installations

Start with the inventory of internet-facing web applications and their deployed plugin versions. When a security release arrives, compare the running version with the vendor’s first fixed build and move affected installations through the change queue ahead of routine upgrades. For PaperCut MF or NG, apply the fixed build for the installed release line, then read the version reported by the running server again. Where web administration is reachable from the internet, restrict it to the management paths that need it while patching; this is Nulltap’s recommended containment for the administrative access route described in the PaperCut record.

For web apps facing new patch releases, preserve reverse-proxy and application access logs across the release window. Look for repeated requests probing newly changed routes, especially followed by unusual outbound transfers or files added under web-served directories; triage such activity against the application’s normal request patterns. GTIG describes forensic traffic and code-level traces from unrestricted autonomous execution, but gives no validated exploit_agent signature or fleet-wide clearance test. Use the logs to investigate suspicious activity, and use the installed version and restricted management reachability to verify the specific exposure changes you made.

Give Web Apps Less Time Exposed

The strongest immediate move is to reduce the interval between a public fix and its deployment on internet-facing apps. Record the running build, the release with the fix, and the time each exposed instance was updated; then check reverse-proxy traffic for abnormal probing during that interval. PaperCut operators have specific fixed-build thresholds in the CVE records. For the broader exploit_agent activity, prioritize exposed software and inspect traffic; a quiet log search alone cannot clear a host with an old, reachable build.

If a newly fixed app stayed reachable during the gap, keep its patch-time logs long enough to investigate the first unexpected requests and any later webshell-like file changes. The staging server GTIG found shows that patch comparison, exploit drafting and execution can run as one automated workflow; defenders still control how long the older version remains available to it.