Insight

Agentic pentest frameworks are no longer a watch-list curiosity: ARTEX, PentAGI and CAIRN appear in Mandiant incident-response engagements, and ARTEX is tied to live espionage

By Riftweld Labs

Insights7 min read
Open-source agentic pentest frameworks are now used in live intrusions
Agentic pentest frameworks are no longer a watch-list curiosity: ARTEX, PentAGI and CAIRN appear in Mandiant incident-response engagements, and ARTEX is tied to

Censys found 1,465 hosts publicly exposing an ARTEX console, while Microsoft tracked 151,000-plus failed logons in a five-day password-spray campaign. Yet the headline framing of the underlying watch-list calls these tools a "second wave" — a contradiction that understates how far agentic offensive frameworks have moved from demonstration into confirmed intrusion activity.

In this article

Why this matters

The operational significance of these frameworks is not that they contain novel exploits. Google Threat Intelligence Group and Mandiant are explicit that the mechanics remain familiar: compromised credentials, unpatched systems, lateral movement. What changed is throughput. In one observed operation, ARTEX produced 27 events in 23 seconds of active burst time while probing AWS permissions across security groups; in another, it trimmed a script after an error and relaunched it to deploy a 17-snapshot instance in roughly 10 seconds. The failed assumption is that human-paced intrusion behavior sets a workable detection ceiling. It no longer does.

That matters because the frameworks are ordinary software. ARTEX is a Go program with an embedded Next.js console; Cairn is described in its own repository as a general-purpose AI state-space search engine validated first on autonomous penetration testing; PentAGI is a multi-agent system for penetration-testing tasks. These are not bespoke implants requiring bespoke signatures. They orchestrate legitimate credentials, legitimate APIs and legitimate administrative functions, which is why Mandiant's defensive recommendations — control-plane isolation, short-lived credentials, least privilege, segmentation and behavioral detection — are long-standing practices rather than new products (Google Threat Intelligence Group).

The second failure of assumption is that repository takedown equals containment. ARTEX's original repository was taken down by its creator in October 2026 because of widespread abuse, yet the framework continued to spread through community forks including ScopeWeaver and a commercial ARTEX PRO. Wiz reported more than 12,000 forks by October 8, 2026. Censys, reported by Infosecurity Magazine, found 1,465 hosts worldwide exposing an ARTEX console, with roughly 815 distinct-looking deployments beyond two identical-hosting blocks; 155 of those hosts, or 10.6 percent, also ran another threat artefact (Wiz; Infosecurity Magazine).

The defensive consequence.
The defensive consequence.

What happened

The strongest evidence is vendor-side incident response. Since late August 2026, GTIG and Mandiant have observed threat actors operating ARTEX, PentAGI, CAIRN and other potentially custom harnesses inside Mandiant engagements, with agents performing significant portions of the attack lifecycle, including lateral movement, database exploitation and exfiltration preparation. The same research describes observed abuse of ARTEX, Hexstrike and CAIRN across multiple phases of intrusion. This is the anchor for the watch-list claim, and it comes from a primary source rather than commentary.

Microsoft's account of Storm-0940, tracked as CovertNetwork-1658 and ORB07, adds concrete impact. The actor used ARTEX in activity dating to September 2026, moved from initial targeting to privileged access in under three hours in multiple cases, and abused a ConnectWise multi-tenant application supply-chain compromise that produced sign-ins to more than 2,250 tenants. Between October 10 and 14, 2026, the actor ran a password-spray campaign from infrastructure at 144.172.64.x: 327 accounts across eight organizations, more than 151,000 failed logons, and a peak of 2,900 failures per minute on October 11 (Microsoft Security Blog).

The financial-sector campaign is corroborated separately. CrowdStrike assessed with moderate confidence that a Chinese-speaking, financially motivated actor used ARTEX against South Korean financial institutions in late September to early October 2026, with DeepSeek v4.1-flash as the primary LLM backend, supplemented by GLM-5.3 and Grok 4.6, accessed through a likely xcai[.]pro API proxy or reseller. Zscaler ThreatLabz confirmed an ARTEX campaign against multiple South Korean financial firms that exfiltrated personally identifiable information relating to approximately 67,000 individuals by exploiting SQL injection and authentication bypass in peripheral internet-facing web applications (CrowdStrike; Zscaler).

Claim Status Basis
ARTEX, PentAGI, CAIRN and custom harnesses seen in Mandiant engagements since late August 2026 Confirmed GTIG/Mandiant primary research
ARTEX used by Storm-0940 in a ConnectWise supply-chain compromise Confirmed Microsoft primary research
ARTEX used against South Korean financial firms, with PII of ~67,000 individuals exfiltrated Confirmed Zscaler and CrowdStrike primary research
1,465 hosts publicly exposing an ARTEX console Reported Censys research via Infosecurity Magazine
CyberStrikeAI drove full Active Directory domain compromise in 3–4 days Reported Kaspersky disclosure via Substack
ARTEX used by "Crafty Falcon" against fintech and crypto Reported CERT-In advisory via Infosecurity Magazine
Any listed package other than ARTEX, PentAGI, CAIRN and Hexstrike appears in the Mandiant incident set Not established Primary source does not address them
Any listed package carries an assigned CVE or formal vendor advisory Not established No such record found in this research
The sequence in the record.
The sequence in the record.

What defenders should do

The practical implication is that detection must be behavioral and volumetric, not signature-based. Mandiant states plainly that static indicators "will not cover much of this activity." The observable signature of agentic use is irregularity: compressed bursts of activity separated by extended pauses, iterative trial-and-error that leaves forensic noise, and API call patterns that no human operator would sustain. Identity telemetry is the highest-value surface, because the observed campaigns repeatedly begin with supplied or compromised credentials and then automate enumeration, privilege escalation and password spraying across tenants.

Three controls deserve priority. First, isolate control planes so an agent that reaches an orchestration console cannot pivot freely into production identity and storage. Second, shorten credential lifetimes and enforce least privilege, so a single supplied credential does not become a path to domain-wide access. Third, alert on velocity rather than on payload: sign-ins per account, failed logons per minute, enumeration breadth and storage access patterns. The ConnectWise case shows why — the supply-chain pivot, not the framework binary, produced the widest tenant reach.

There is also a detection opportunity in infrastructure overlap. Censys found ARTEX sharing infrastructure with CyberStrikeAI, PentAGI, HexStrike AI and RedAmon, with CyberStrikeAI alongside ARTEX on 70 hosts. That correlation is a lead-generation signal, not a complete detection strategy, and it should be treated as such.

Limits of the record

The watch-list itself is a secondary compilation. RuntimeWire's article enumerates 18 packages — ARTEX, Cairn, CyberStrike, PentAGI, CyberStrikeAI, PentestGPT, PentestAgent, AutoPentester, VulnBot, RedAmon, MAPTA, Strix, Shannon, BugTraceAI, Nebula, hackingBuddyGPT, HexStrike AI and CAI — and adds a second wave of open-source pentest agents to those already covered in incident reporting. The article's own headline and body say 18, while the URL slug and the framing of the seed post say 17; the enumerated list contains 18 names. RuntimeWire also notes that CyberStrike was not named in the Google/Mandiant report (RuntimeWire).

Beyond ARTEX, PentAGI, CAIRN and Hexstrike, the record does not establish live operational use of the remaining entries. The Kaspersky CyberStrikeAI disclosure, the NorForge-observed Cairn panels, the Seoul extortion-ring account and the CERT-In "Crafty Falcon" advisory all rest on single secondary sources, and the underlying CERT-In advisory was not directly retrieved. The Shanholo post recommending CyberStrike, Cairn and PentAGI is a recommendation, not an incident report. The Reddit seed post was removed by moderators before it could be read directly, so it adds no independent evidence.

The record is also thin on attribution and on formal vulnerability handling. CrowdStrike's assessment of the South Korean financial actor is explicitly moderate confidence. No CVE or vendor advisory was found for any listed package, which means defenders cannot treat these frameworks as patchable products with versioned fixes; they are moving targets distributed through forks.

Method and sources

This analysis draws on primary research published by Google Threat Intelligence Group and Mandiant, Microsoft, Wiz, CrowdStrike and Zscaler, and on secondary reporting from RuntimeWire, Infosecurity Magazine, Cybersecurity Dive and a Substack account citing Kaspersky and NorForge. Repository links for ARTEX, Cairn, CyberStrike, PentAGI, CyberStrikeAI, HexStrike AI and CAI were used to confirm project identity and stated function, not to verify incident involvement.

Claims are labeled by evidentiary strength: Confirmed for primary-source findings, Reported for secondary sourcing, Observed for directly described telemetry, Inferred for conclusions drawn from those records, and Not established where the record is silent. The central claim — that open-source agentic offensive frameworks are in confirmed live use — rests on multiple independent primary sources. The broader claim that all 18 listed packages warrant equal operational attention is not established by the available reporting.

Adversary simulation

Prove the route before someone else does.

We reproduce the path an attacker would take, show the impact with evidence, then retest after the fix.