Threat reportThreat IntelligenceTL-2026-2793
Agentic AI used for post-exploitation in breach of the Dutch Institute for Vulnerability Disclosure (DIVD)
The Dutch Institute for Vulnerability Disclosure (DIVD), a volunteer-run security nonprofit, disclosed a breach in which an unidentified attacker exploited a technical vulnerability in an undisclosed system and then used an autonomous agentic AI for post-exploitation, including password spraying and an adversary-in-the-middle operation.
Key facts
| Identifiers | TL-2026-2793 |
|---|---|
| Severity | HIGH |
| Affected products | Dutch Institute for Vulnerability Disclosure (DIVD) Undisclosed system with an undisclosed technical vulnerability (DIVD states not Citrix NetScaler) |
| Tools | Autonomous agentic AI post-exploitation agent (unidentified model/framework) |
| MITRE ATT&CK techniques | T1059 Command and Scripting Interpreter, T1110.003 Password Spraying, T1190 Exploit Public-Facing Application, T1557 Adversary-in-the-Middle, T1588.007 Obtain Capabilities: Artificial Intelligence |
| Motivation | UNKNOWN |
| Target sectors | cybersecurity, nonprofit, research |
| Target regions | netherlands, Europe |
| Detection rules | 9 (1 shown free) |
| Sources | 6 cited sources |
Detection coverage
As of 2026-09-29, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-2793 across Splunk SPL, Microsoft KQL and Sigma, covering 7 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: DIVD Agentic Intrusion — Web/App Server Process Spawning Shell or Script Interpreter
Detects an internet-facing web or application-runtime process (IIS w3wp, Apache, nginx, PHP-CGI, Node, Java/Tomcat) spawning a command shell or script interpreter. DIVD reports initial access through an undisclosed technical vulnerability in a public-facing system (not Citrix NetScaler; no CVE published), so the exploit itself cannot be signatured. The reliable, product-agnostic signal is the execution that follows exploitation of any public-facing service, which is followed by the agent's adaptive post-exploitation.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1190, T1059
- Data sources: sysmon, microsoft_defender_for_endpoint
- Severity: high
- Confidence: medium
Splunk SPL
```Web/app-server process spawning a shell or script interpreter (T1190 -> T1059)```
index=sysmon sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1
| regex ParentImage="(?i)(^|[^a-z0-9_])(w3wp|httpd|nginx|php-cgi|node|java|tomcat\d*)\.exe$"
| regex Image="(?i)(^|[^a-z0-9_])(cmd|powershell|pwsh|wscript|cscript|mshta|python|certutil|bitsadmin|curl)\.exe$"
| stats count min(_time) as first_seen max(_time) as last_seen values(CommandLine) as command_lines values(Image) as child_images by host, ParentImage, User
| convert ctime(first_seen) ctime(last_seen)Microsoft KQL
// Web/app-server process spawning a shell or script interpreter (Windows and Linux MDE telemetry)
DeviceProcessEvents
| where Timestamp > ago(24h)
| where InitiatingProcessFileName in~ ("w3wp.exe","httpd.exe","nginx.exe","php-cgi.exe","node.exe","java.exe","tomcat9.exe","apache2","httpd","nginx","php-fpm","java","node")
| where FileName in~ ("cmd.exe","powershell.exe","pwsh.exe","wscript.exe","cscript.exe","mshta.exe","python.exe","certutil.exe","bitsadmin.exe","curl.exe","sh","bash","dash","python","python3","perl","curl","wget","nc")
| summarize Executions=count(), FirstSeen=min(Timestamp), LastSeen=max(Timestamp), CommandLines=make_set(ProcessCommandLine, 10) by DeviceName, AccountName, InitiatingProcessFileName, FileName
| project DeviceName, AccountName, InitiatingProcessFileName, FileName, Executions, FirstSeen, LastSeen, CommandLinesSigma
title: Web Server or Application Runtime Spawning Shell or Script Interpreter
id: 6f1c2a4e-3b7d-4c58-9a21-5e0d8b7c1a01
status: experimental
description: Detects an internet-facing web or application-runtime process spawning a command shell or script interpreter, the product-agnostic post-exploitation signal after exploitation of a public-facing application (DIVD agentic-AI breach, undisclosed vulnerability).
references:
- https://csirt.divd.nl/2026/09/24/when-not-if/
- https://www.bleepingcomputer.com/news/security/automated-ai-agent-used-to-breach-cybersecurity-nonprofit-divd/
author: Threadlinqs Team
date: 2026/09/30
tags:
- attack.initial_access
- attack.t1190
- attack.execution
- attack.t1059
logsource:
category: process_creation
product: windows
detection:
selection_parent:
ParentImage|endswith:
- '\w3wp.exe'
- '\httpd.exe'
- '\nginx.exe'
- '\php-cgi.exe'
- '\tomcat9.exe'
- '\java.exe'
- '\node.exe'
selection_child:
Image|endswith:
- '\cmd.exe'
- '\powershell.exe'
- '\pwsh.exe'
- '\wscript.exe'
- '\cscript.exe'
- '\mshta.exe'
- '\python.exe'
- '\certutil.exe'
- '\bitsadmin.exe'
- '\curl.exe'
condition: selection_parent and selection_child
falsepositives:
- Web applications that legitimately shell out (report generators, deployment hooks, CI agents running on Java or Node)
- Administrative maintenance scripts launched by an application pool
level: highKnown false positives:
- Web applications that legitimately shell out to helper scripts (report generation, image conversion, deployment hooks)
- Java/Node based CI or build agents that spawn cmd/powershell as part of normal jobs
Tuning notes: Allowlist known (ParentImage, CommandLine) pairs per host role and exclude build servers. Extend the parent list to whichever internet-facing products you run; add Linux auditd/Sysmon-for-Linux parents if applicable. Because DIVD has not named the vulnerable product, keep this rule product-agnostic and re-scope once the 2026-10-01 DIVD update identifies it.
- DIVD Agentic Intrusion — Entra ID Password Spray Followed by Successful Sign-in · unlock with TL Creds
- DIVD Agentic Intrusion — PowerShell Script Block With Dense Explanatory Comments and Recon/Spray Primitives · unlock with TL Creds
- DIVD Agentic Intrusion — Rapid Burst of Distinct Discovery Utilities From One Parent · unlock with TL Creds
- DIVD Agentic Intrusion — Non-Browser Process Resolving LLM API Endpoints (Possible On-Host Agent) · unlock with TL Creds
- DIVD Agentic Intrusion — ARP Poisoning Indicator: Address Conflict Claimed by One MAC Across Hosts · unlock with TL Creds
- DIVD Agentic Intrusion — Password-Spray and AiTM Tool Execution · unlock with TL Creds
- DIVD Agentic Intrusion — Windows Password Spraying From a Single Source · unlock with TL Creds
- DIVD Agentic Intrusion — LLMNR/NBT-NS/mDNS Poisoning Responder (AiTM) · unlock with TL Creds
Indicators of compromise
As of 2026-09-29, Threadlinqs Intelligence tracks 7 indicators of compromise for TL-2026-2793.
Technical analysis
DIVD called the intrusion 'loud and very, very messy'. The scope of compromised data is not yet established.
On 2026-09-24 (23:34 CEST) DIVD's CSIRT published 'It was a matter of when, not if…', disclosing that it had noticed suspicious activity in its infrastructure, concluded a breach had occurred, blocked access to its infrastructure, and engaged an external incident response team for forensics. DIVD said this was the first breach it had suffered in nearly seven years of operation, and that it was treating the incident as a worst-case scenario, assuming compromise until proven otherwise.
Initial access: according to DIVD (as reported by BleepingComputer, DataBreaches.net and Beinsure), the actor exploited a technical vulnerability in an undisclosed system. DIVD stated the vulnerability is not Citrix NetScaler. No CVE, product name or vendor has been published; DIVD is withholding details to avoid influencing the investigation and to protect other potentially affected victims, which it plans to notify.
Post-exploitation: DIVD stated 'This is an attack we have not seen before. Not because it's our first, but because the modus operandi indicates that this is an agentic AI-powered attack.' Evidence indicates the agent chose its next action after each previous step rather than following a pre-scripted sequence. It operated quickly but with sloppy logic, performed password spraying while simultaneously running an adversary-in-the-middle operation (thereby interfering with its own AiTM activity), and left comments explaining many of its actions in the code it produced. The intrusion was unusually noisy and left substantial forensic evidence. DIVD suggested the agent appeared poorly trained or configured for offensive security work.
Response and status: DIVD reported the incident to the Autoriteit Persoonsgegevens (Dutch data protection authority) and the NCSC, and consulted/informed the police. Directly affected parties were notified. Whether data was accessed, stolen or altered is unconfirmed. The actor's identity, the vulnerable product and the exact affected systems are undisclosed. No IOCs (IPs, domains, hashes) were published. DIVD's initial notice scheduled an update for 2026-09-28; reporting on 2026-09-29 cites a further detailed update promised for 2026-10-01. Severity is an analyst estimate pending confirmation of impact.
MITRE ATT&CK techniques by tactic:
Execution
T1059 Command and Scripting Interpreter
Credential Access
T1110.003 Password Spraying; T1557 Adversary-in-the-Middle
Initial Access
T1190 Exploit Public-Facing Application
Resource Development
Remediation
Patches
- No CVE or vendor patch identified; DIVD states the flaw is not Citrix NetScaler. Monitor the DIVD CSIRT update promised for 2026-10-01
Immediate actions
- Hunt authentication logs for low-and-slow and high-volume password spraying against VPN, SSO/IdP, OWA and admin portals, and enforce lockout/throttling and MFA
- Review internal network segments for rogue ARP/DHCP/LLMNR/NBT-NS/WPAD responders and other adversary-in-the-middle activity
- Look for scripts or tool output containing verbose, explanatory natural-language comments dropped on hosts, which DIVD reported as an artifact of the agent
- Rotate credentials for accounts exposed to the affected infrastructure and apply patches for internet-facing systems as vendor fixes emerge
Workarounds
- Isolate or restrict access to affected internet-facing systems until the vulnerable product is identified
- Block or rate-limit repeated authentication failures per source and per account
Longer-term hardening
- Reduce internet-facing attack surface and keep an accurate inventory of exposed services
- Deploy behavioral EDR/NDR that detects rapid, sequential, adaptive post-exploitation rather than fixed tooling signatures
- Enforce phishing-resistant MFA and conditional access to blunt credential spraying and AiTM credential capture
- Prepare an open, transparent incident-communication process modeled on DIVD's disclosure
Affected and fixed versions
- Dutch Institute for Vulnerability Disclosure (DIVD) — Undisclosed system with an undisclosed technical vulnerability (DIVD states not Citrix NetScaler)
Timeline
- DIVD established as a Dutch volunteer-run coordinated vulnerability disclosure nonprofit; DIVD says the 2026 incident is its first breach in nearly seven years of operation
- DIVD CSIRT publishes 'It was a matter of when, not if…' (23:34 CEST), stating the modus operandi indicates an agentic AI-powered attack and that the incident was reported to the Autoriteit Persoonsgegevens, NCSC and police
- DIVD notices suspicious activity in its infrastructure, concludes a breach occurred, blocks access to infrastructure and engages an external incident response team
- DataBreaches.net reports DIVD is investigating an agentic AI-powered attack; DIVD treating it as a worst-case scenario
- Scheduled DIVD update date; DIVD provides an incident update describing the agent's behavior (next-step decisions, password spraying interfering with its own AiTM, explanatory code comments) and that the flaw is not Citrix NetScaler
- BleepingComputer publishes details of the automated AI agent used to breach DIVD; scope of data compromise still undetermined
- DIVD promises a detailed update, including disclosures about other potential victims of the same vulnerability
Sources cited
- It was a matter of when, not if… (DIVD CSIRT) csirt.divd.nl · Advisory · 2026-09-24
- Dutch Institute for Vulnerability Disclosure (Wikipedia) en.wikipedia.org · Research · 2026-09-29
- DIVD reports suspected agentic-AI breach (ai-hack-watch issue #36) GitHub · Code repository · 2026-09-28
- Automated AI agent used to breach cybersecurity nonprofit DIVD (BleepingComputer) BleepingComputer · News · 2026-09-29
- DIVD Dutch Institute for Vulnerability Disclosure investigating agentic AI-powered attack (DataBreaches.Net) databreaches.net · News · 2026-09-25
- DIVD says AI agent carried out cyberattack through software flaw (Beinsure) beinsure.com · News · 2026-09-29
Previous and next in threat intel
Drafted by the Threadlinqs research pipeline from 6 sources; published ; last updated —.