# exploit_agent: AI-driven patch-diffing framework autonomously discovers, weaponizes and executes web application exploits (GTIG)

> Google Threat Intelligence Group (GTIG) found an exposed staging server hosting 'exploit_agent', a multi-agent framework that uses frontier and local LLMs to diff web application software before and after patch announcements, draft exploits and execute them with no human in the loop. GTIG calls this the first time it has seen vulnerability discovery and exploitation fully automated this way. The operator showed Russian-language indicators and appeared interested in French media/entertainment websites and the education sector.

- **Published:** 2026-11-16T14:14:17Z
- **Last reviewed:** 2026-11-16T14:14:17Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-6150
- **Card image:** https://intel.threadlinqs.com/og/threat/TL-2026-6150.png
- **ID:** TL-2026-6150
- **Severity:** HIGH
- **Category:** THREAT_INTEL
- **Status:** ACTIVE
- **Detections:** 9 (1 shown free below) · **IOCs:** 10
- **CVEs:** CVE-2026-7465

## Description

On 2026-11-16 GTIG published its AI Threat Tracker ('Agentic threat activity crosses a significant threshold'). It reports an exposed staging server that hosted exploit_agent, a multi-agent vulnerability research and weaponization framework integrated with multiple frontier and local LLMs. The framework is configured to run continuous differential analysis of popular web application software. It compares code before and after patch announcements, analyzes the differences to find potential vulnerabilities, drafts exploits and executes them automatically. GTIG states this is the first time it has identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop.

The staging server also held data exfiltrated from victim organizations and several post-compromise scripts. These covered internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells. A configuration file carried the Russian comment '# LLM API-ключи для авто-эксплуатации (пилот exploit_agent)' ('LLM API keys for auto-exploitation (exploit_agent pilot)'), with further Russian instructions to enter the keys directly. Victim telemetry suggests the operator was interested in French websites in the media and entertainment industries and in the education sector. The sources name no victims, victim counts, CVEs, IP addresses, domains or file hashes, and no named actor. Attribution is limited to a suspected Russian-speaking operator, so confidence is low.

As an example of the framework's output, GTIG shows it inspecting access control in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin v2.19.25. The figure shows the restrict_popup_rest_access function in v21925.php, which guards the REST route /wp/v2/spectra-popup, flagged as an insecure capability check. GTIG gives no CVE for this and does not say it was exploited in the wild. A separate Wordfence entry (CVE-2026-7465, authenticated Contributor+ RCE in Spectra <= 2.19.25) was not tied to exploit_agent by any source reviewed, so it is not listed as this threat's CVE.

GTIG places the finding in a wider shift to agentic offensive tooling. Other frameworks it names are ARTEX (Chinese-developed, open source, released July 2026), CAIRN (OODA-loop multi-agent pentesting, seen in intrusion activity including against US aerospace), CyberStrikeAI (Go-based C2 and orchestration for 100+ tools), HexStrike AI (MCP harness with 150+ tools) and PentAGI (a customized build used from late August to early October 2026 for reconnaissance and vulnerability scanning against organizations in Asia). PRC-nexus espionage actors are also experimenting with such frameworks. Press coverage describes a Chinese-speaking operator chaining Hermes, Strix and Cairn against 27+ companies in September 2026, and the use of ARTEX against South Korean financial organizations and against 123 Reg. These are context for the trend, not confirmed links to the exploit_agent operator.

Defensive implication: the window between a patch being published and exploitation shrinks because n-day exploits can be built by diffing patches automatically. Priorities are fast patching of internet-facing web applications (CMS plugins especially), webshell detection, rotation of credentials, and monitoring for unusual LLM service enablement and for 'burn and rotate' service account patterns.

## MITRE ATT&CK

- T1595.002 Vulnerability Scanning
- T1583.004 Server
- T1587.004 Exploits
- T1588.006 Vulnerabilities
- T1588.007 Artificial Intelligence
- T1190 Exploit Public-Facing Application
- T1059 Command and Scripting Interpreter
- T1505.003 Web Shell
- T1110 Brute Force

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**exploit_agent — Spectra Popup REST Route Probing and Write Requests** (`TL-2026-6150-SPL-001`)

Behaviour-first search (TL-2026-6150) for automated probing or write-method requests to the Spectra (Ultimate Addons for Gutenberg) REST route /wp/v2/spectra-popup, the access-control target (restrict_popup_rest_access, v21925.php) GTIG showed exploit_agent analysing. Flags request bursts or successful writes per source and 15-minute window. The route is an analysis target, not a confirmed exploitation indicator.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1190, T1595.002
- **Data sources:** splunk:index=web:access_combined, splunk:index=web:ms:iis:auto, splunk:index=web:nginx:plus:kv, W3CIISLog
- **Severity:** high
- **Confidence:** medium

### Splunk SPL

````spl
index=web sourcetype IN ("access_combined","ms:iis:auto","nginx:plus:kv") "spectra-popup"
```Spectra REST route wp/v2/spectra-popup guarded by restrict_popup_rest_access in v21925.php was analysed by exploit_agent```
| eval src=coalesce(clientip,c_ip,src_ip), meth=upper(coalesce(method,cs_method)), code=tostring(coalesce(status,sc_status))
| eval is_write=if(match(meth,"^(POST|PUT|PATCH|DELETE)$"),1,0), is_ok_write=if(is_write=1 AND match(code,"^(200|201|204)$"),1,0)
| bin _time span=15m
| stats count AS requests sum(is_write) AS write_requests sum(is_ok_write) AS successful_writes values(meth) AS methods values(code) AS statuses by _time, host, src
| where requests>=10 OR successful_writes>=1
| table _time host src requests write_requests successful_writes methods statuses
````

### Microsoft KQL

```kql
W3CIISLog
| where TimeGenerated > ago(24h)
| where csUriStem has "spectra-popup" or csUriQuery has "spectra-popup"
| extend IsWrite = csMethod in~ ("POST", "PUT", "PATCH", "DELETE")
| summarize Requests = count(), WriteRequests = countif(IsWrite), SuccessfulWrites = countif(IsWrite and scStatus in ("200", "201", "204")), Methods = make_set(csMethod), Statuses = make_set(scStatus), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated) by cIP, sSiteName, bin(TimeGenerated, 15m)
| where Requests >= 10 or SuccessfulWrites >= 1
| project TimeGenerated, sSiteName, cIP, Requests, WriteRequests, SuccessfulWrites, Methods, Statuses, FirstSeen, LastSeen
```

### Sigma

```yaml
title: Write-Method Request to Spectra Popup REST Route (exploit_agent Target)
id: 3f6a2c1e-8b74-4d09-a5e2-61c0b7d94a11
status: experimental
description: Detects POST, PUT, PATCH or DELETE requests to the WordPress REST route /wp/v2/spectra-popup of the Spectra (Ultimate Addons for Gutenberg) plugin, whose restrict_popup_rest_access check in v21925.php was analysed by exploit_agent (TL-2026-6150). This is an access-control analysis target and not a confirmed exploitation indicator.
references:
  - https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
  - https://www.theregister.com/research/2026/11/16/google-ai-agent-turned-a-security-patch-into-a-working-exploit-with-no-human-help/5309926
author: Threadlinqs Team
date: 2026-11-16
tags:
  - attack.initial-access
  - attack.t1190
  - attack.reconnaissance
  - attack.t1595.002
logsource:
  category: webserver
detection:
  selection_route:
    c-uri|contains: 'spectra-popup'
  selection_method:
    cs-method:
      - 'POST'
      - 'PUT'
      - 'PATCH'
      - 'DELETE'
  condition: selection_route and selection_method
falsepositives:
  - Authenticated editors saving popup content through the WordPress block editor
  - Authorised vulnerability scanning of the site
level: medium

```

**Known false positives:**

- Authenticated editors saving popups through the block editor generate write requests
- Authorised vulnerability scanners or uptime monitors requesting the route repeatedly

**Tuning notes:** Exclude authenticated admin/editor source ranges and known scanner IPs; lower the request threshold (10 per 15 min) on low-traffic sites; adjust the index and sourcetype to your web log pipeline; combine with WAF logs to separate unauthenticated requests. Expect the route to be probed by many scanners after any Spectra advisory.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- exploit_agent — Automated Brute Force Against WordPress Login and XML-RPC (locked: unlock with TL Creds)
- exploit_agent — Web Server Process Spawning Shell or Recon Utility (locked: unlock with TL Creds)
- exploit_agent — Agentic Offensive Framework Execution (locked: unlock with TL Creds)
- exploit_agent — Webshell Access via PHP in WordPress Uploads Directory (locked: unlock with TL Creds)
- exploit_agent — Brute-Force and Recon Tools Launched From Web Server Context (locked: unlock with TL Creds)
- exploit_agent — Unexpected Local LLM Inference Service on Server (locked: unlock with TL Creds)
- exploit_agent — Webshell Written or Rival Webshell Removed by Web Server Process (locked: unlock with TL Creds)
- exploit_agent — Server-Side Process Contacting LLM Provider APIs (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sources

- [GTIG AI Threat Tracker: Agentic threat activity crosses a significant threshold](https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold)
- [Google: AI agent turned a security patch into a working exploit with no human help (The Register)](https://www.theregister.com/research/2026/11/16/google-ai-agent-turned-a-security-patch-into-a-working-exploit-with-no-human-help/5309926)
- [Crook used three open-source agents to break into a Fortune 500 hospitality company, a major US airline and 25 other orgs (The Register)](https://www.theregister.com/security/2026/09/25/crook-used-three-open-source-agents-to-break-into-a-fortune-500-hospitality-company-a-major-us-airline-and-25-other-orgs/5299012)
- [123 Reg discloses AI cyberattack on customer data, suspects China nexus (The Register)](https://www.theregister.com/cyber-crime/2026/10/28/123-reg-discloses-ai-cyberattack-on-customer-data-suspects-china-nexus/5305936)
- [Threat actor targets Indian fintechs on weekends when defenders aren't watching (The Register, CyberStrikeAI)](https://www.theregister.com/security/2026/11/06/threat-actor-targets-indian-fintechs-on-weekends-when-defenders-arent-watching/5308000)
- [Crims boast of using HexStrike AI against Citrix bugs (The Register)](https://www.theregister.com/security/2025/09/03/crims-boast-of-using-hexstrike-ai-against-citrix-bugs/368681)
- [Wordfence: Spectra Gutenberg Blocks 2.19.25 Authenticated Contributor RCE (CVE-2026-7465), context only, not linked to exploit_agent by sources](https://www.wordfence.com/threat-intel/vulnerabilities/wordpress-plugins/ultimate-addons-for-gutenberg/spectra-gutenberg-blocks-21925-authenticated-contributor-remote-code-execution-via-arbitrary-php-function-call-via-block-attributes)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-6150
