On 7 October 2026, CrowdStrike published a report on a campaign targeting the South Korean financial sector. In roughly two weeks, seven financial institutions were compromised and the personal data of several tens of thousands of customers was exfiltrated. The attacker was not a team of seasoned hackers working around the clock: it was a single operator, backed by an AI-driven autonomous pentesting tool called ARTEX.
A month later, on 16 November, TechCrunch reported that an FBI alert describes ARTEX-linked attacks going from target to active compromise in under three hours.
ARTEX is not some obscure piece of malware. It is an open source project under the AGPL-3.0 licence, and its “global edition” translated into five languages, jiwoochris/artex-global, has more than 3,700 stars and 3,200 forks on GitHub at the time of writing.
In this article, we break down what autonomous AI pentesting is, how ARTEX works under the hood, what the Korean incident teaches us, and above all what you should actually do so that your website, back office or API does not become the next easy target.
Disclaimer: this article is defensive and educational. Scanning, probing or exploiting a system without its owner’s written permission is a criminal offence in France (article 323-1 of the Criminal Code) and in most other countries. We provide no offensive procedure here.
Autonomous AI pentesting: what are we talking about?
From the classic scanner to the autonomous agent
A traditional vulnerability scanner runs a fixed list of checks, in a fixed order, then stops. It does not “understand” anything: it compares responses against known signatures.
An autonomous pentesting agent works differently. A large language model (LLM) receives a goal, for example “gain access to acme.com’s customer database”, and:
- breaks that goal down into sub-tasks on its own;
- runs real tools (HTTP requests, shell scripts, reconnaissance tools);
- reads the results, draws conclusions and chooses the next step;
- starts again until it reaches the goal or runs out of leads.
The difference is fundamental: requests are no longer hard-coded constants, but values the model generates from context. Each attempt looks a little less like the previous one.
A lasting trend, not an isolated case
ARTEX is part of a broader movement. As early as 2024, researchers at the University of Illinois showed that a GPT-4-based agent could exploit 87% of a sample of one-day vulnerabilities from their CVE description alone. In 2025, the commercial agent XBOW climbed to the top of the US leaderboard on the HackerOne bug bounty platform.
On the threat side, in November 2025 Anthropic documented the first espionage campaign largely orchestrated by AI: a group attributed to a Chinese state actor had turned Claude Code against around thirty targets. According to Anthropic, the AI carried out 80 to 90% of the operational work, peaking at several requests per second, a pace no human team could match.
In February 2026, the ANSSI threat summary on generative AI (the French national cybersecurity agency) still considered that no AI system had managed to carry out every stage of an attack on its own. Eight months later, the ARTEX case shows how fast that line is moving: a human remains in control, but most of the attack chain is now automated.
ARTEX under the hood: architecture of an offensive agent
Understanding ARTEX’s architecture is far from academic: it explains why traditional defences struggle to spot it, and which signals it still leaves behind.
A modern, compact stack
ARTEX is a single binary written in Go, which embeds a Next.js interface and stores its data in PostgreSQL. Its agent capabilities rely on the norma SDK. The original project, Autumn-27/ARTEX, was documented in Chinese; it is no longer accessible. jiwoochris’s global edition translates the interface and reports into English, Korean, Chinese, Spanish and Japanese.
The LLM is interchangeable: Anthropic models, OpenAI models, or any OpenAI-compatible endpoint (open source models or models hosted outside the United States). According to CrowdStrike, quoted by Hunt.io, the Korean attacker mainly used DeepSeek, with GLM (Zhipu AI) and Grok as fallbacks.
Four specialised agent roles
The work is split between several agents with distinct roles:
- goals: breaks down the initial objective and extracts the scope;
- planner: the only agent that decides which direction to explore next (the “intents”);
- workers: each takes a single intent, executes it with real tools, records its results and stops (three in parallel by default);
- mainagent: the human-in-the-loop entry point, through a chat interface.
The dual graph: the attack’s memory
This is the core of the system. ARTEX separates two questions:
- The asset graph answers “what exists?”: domains, subdomains, IPs, services, endpoints. It is shared across all tasks.
- The exploration graph answers “how far have we tested?”: goals, intents, observed facts, confirmed vulnerabilities.
The two are linked by anchors. The result: the attack deepens incrementally. An observation made on a subdomain on Tuesday feeds a new lead on Wednesday, and is never forgotten.
A closed loop that never gets tired
The engine is event-driven. Every time the graph changes, the planner wakes up, assesses the situation and may hand out new intents. The workers write their results, which changes the graph… and restarts the cycle.
For sequential attack chains (find an injection → obtain credentials → move laterally → escalate privileges), the planner keeps a persistent to-do list. It only starts the next step once the previous one has produced the expected fact. This is what lets ARTEX carry multi-stage attacks through to the end without derailing.
Workers can even search the execution traces of other workers: an unusual error message or a hidden parameter spotted by one immediately benefits the next ones.
Key takeaway: a classic scanner is a gun firing a burst. An agent like ARTEX is a tireless investigator who takes notes, cross-checks clues and comes back through another door.
The Korean case: what the incident teaches us
Timeline and impact
Based on information published by CrowdStrike, Genians, Hunt.io and TechCrunch:
- late September to early October 2026: seven South Korean financial institutions suffer data leaks through auxiliary systems (a loan simulation service, an internal employee tool, credit brokers);
- impact: around 66,000 to 68,000 people affected depending on the source, plus about 2,200 companies according to Hunt.io;
- 7 October: CrowdStrike attributes the campaign, with moderate confidence, to a single Chinese-speaking, financially motivated operator;
- 11 October: according to Genians, the original GitHub repository and the official Docker image are taken down, but the forks remain online;
- mid-November: the FBI issues an alert. It links ARTEX to attacks against finance, government, tech, defence, telecoms and law firms. According to TechCrunch, the developer has since closed the code, but the former open source version remains “widely accessible”.
Three very practical lessons
1. The attackers did not need a zero-day. The FBI points to weak protections, exposed or stolen passwords and the absence of multi-factor authentication. AI does not invent new flaws: it exploits the ones you have not fixed, faster and more systematically.
2. The back doors are the “side” systems. It was not the core banking applications that gave way, but peripheral tools: simulators, partner portals, internal tools. For a website, the equivalent is the forgotten staging site, the old recruitment form or the plugin installed “just to test” back in 2022.
3. Attackers expose their tools too. Researchers found hundreds of ARTEX instances publicly reachable without a password, including the one used against Korea. An ironic reminder: default configuration matters as much as the tool itself.
How to detect an autonomous pentesting agent
The good news: ARTEX’s global edition publishes a very thorough defence and detection guide, with ready-to-use rules. Its core principle is worth remembering: static fingerprints change, behaviour is hard to hide.
Static fingerprints (supporting clues)
On the target side, two User-Agents specific to the ARTEX ecosystem may show up in your logs:
artex-enrich/1.0: the automatic enrichment module, which sends a simpleGET, reads the page’s<title>and closes the connection;norma/0.4: the WebFetch tool of the norma SDK (not exclusive to ARTEX).
A quick search through your Nginx or Apache logs is enough for a first check:
# Search access logs for ARTEX-related User-Agents
zgrep -hE 'artex-enrich/|norma/' /var/log/nginx/access.log* \
| awk '{print $1}' | sort | uniq -c | sort -rn | head
Common mistake: concluding “nothing found, so nothing to worry about”. The operator can change the User-Agent in one line, and the tools launched by the workers (curl, nmap…) use their own signatures. No clue is not proof of safety.
Behavioural signatures (the real signal)
What gives an autonomous agent away is the co-occurrence of several behaviours from the same source (or a small group of rotating sources):
- reconnaissance → directory enumeration → parameter testing → authentication or injection attempts, chained at short intervals;
- successive requests on the same endpoint that mutate consistently depending on the HTTP status code and response size (adaptive, not random);
- a session that keeps going without a break for hours, including at night;
- persistence after failure: 401, 403 or 429 responses that do not discourage it, but trigger bypass variants.
Sigma and Suricata rules ready to deploy
The repository’s detections/ folder contains:
- Sigma rules (atomic detection and correlations) that can be converted to Splunk, Elasticsearch, Microsoft Sentinel, etc.;
- Suricata rules for the network layer;
- a MITRE ATT&CK Navigator layer showing which technique each rule covers;
- a list of indicators in CSV and MISP formats;
- a read-only, dependency-free Python host triage script.
To convert the rules for your SIEM, the repository documentation says:
python3 -m venv .venv && . .venv/bin/activate
pip install sigma-cli
# Example: convert all rules to Splunk
sigma convert -t splunk --without-pipeline detections/sigma/
The guide also offers a generic correlation model (enumeration + probing + authentication from the same IP within 10 minutes) that does not rely on any ARTEX signature. It is probably the most useful rule to adapt, because it also covers future tools of the same kind.
Pro tip: behind a CDN or a reverse proxy, group by the client’s real IP (validated
X-Forwarded-For) or by session ID, otherwise all your correlations will land on the proxy’s IP.
No SIEM? Accessible alternatives
Not every small business has a SOC. A few tools make behavioural detection possible without an outsized budget:
- CrowdSec: an open source engine from France that analyses logs against behavioural scenarios (scanning, brute force, etc.) and blocks through “bouncers”, including one for WordPress;
- Wazuh: an open source SIEM/XDR that can ingest adapted rules;
- Wordfence: a web application firewall and two-factor authentication for WordPress.
Hardening your website against AI-driven attacks
Detection is good. Giving attackers nothing to grab is better. The recommendations below build on ARTEX’s defence guide and on the FBI’s findings, adapted to the reality of a business website.
1. Reduce the attack surface
An autonomous agent starts by mapping everything that is exposed. Do it before it does:
- inventory your subdomains, staging environments, admin consoles and APIs;
- take off the internet everything that does not need to be there (VPN, IP allowlist, HTTP basic authentication on staging);
- delete unused plugins, themes and tools rather than just deactivating them.
2. Patch known vulnerabilities quickly
The priority target is not the brand-new flaw, it is the published, unpatched vulnerability. To prioritise, rely on the CISA KEV catalog (actively exploited vulnerabilities) and on advisories from CERT-FR or your national CERT.
On WordPress, a quick status check with WP-CLI:
# List plugins and themes with an available update
wp plugin list --update=available --fields=name,version,update_version
wp theme list --update=available --fields=name,version,update_version
wp core check-update
3. Enforce MFA, no exceptions
This point comes up in every report. Turn on multi-factor authentication for:
- every administrator and editor account on your CMS;
- your hosting, DNS, registrar and CDN consoles;
- your business SaaS tools (CRM, invoicing, email).
Add adaptive rate limiting to login and password reset pages, and check your credentials against known breaches, for example with Have I Been Pwned.
4. Treat side systems like the main site
Quote forms, partner areas, candidate portals, webhooks: these secondary entry points are often less protected. They need the same level of authentication, updates and monitoring as your main application.
For APIs, check server-side that the user is actually allowed to access each requested object (an IDOR flaw). An AI agent mechanically increments identifiers and tests everything, at scale. The same applies to the endpoints you deliberately open to agents, such as those described in our guide to WordPress, the Abilities API and MCP (in French).
5. Leak nothing in error messages
An autonomous agent reads your error messages carefully to steer its next requests. Turn off error display in production (WP_DEBUG_DISPLAY set to false on WordPress), and never return a stack trace, server path or API key in a response.
6. Keep logs and prepare the response
An attack that succeeds in three hours leaves no time to look for where the logs are. Make sure you keep access, authentication and outbound network logs for long enough.
Also prepare automatic responses (temporarily blocking a source, invalidating sessions): waiting for human approval of every measure is incompatible with an agent’s speed.
Common mistake: copying a list of “malicious IPs” found on a forum into your firewall. ARTEX’s guide explicitly advises against it: these lists go stale within hours (agents change IPs constantly) and may block legitimate customers. Prefer indicators from official sources or verified threat intelligence feeds.
Using these tools for defence: legal framework and good practice
Autonomous pentesting has legitimate uses too
Security teams already use AI agents to test their own systems more often and more broadly. ARTEX’s architecture (dual graph, full request traceability through a recording proxy, approval queue for destructive commands) shows what a serious professional tool should offer.
If you want to study this kind of tool, do it only in an isolated lab, against intentionally vulnerable targets such as OWASP Juice Shop, and keep the interface bound to 127.0.0.1 (artex-global’s default behaviour). The hundreds of instances exposed without a password are the perfect counter-example. The same rule applies to any self-hosted agent, such as an OpenClaw assistant installed on a VPS (in French): never leave an admin interface open to the internet without authentication.
What French law says
- Article 323-1 of the Criminal Code: fraudulently accessing or remaining in an automated data processing system is punishable by 3 years’ imprisonment and a €100,000 fine, more if data is altered.
- Article 323-3-1: holding or making available a tool designed for these offences is also punishable, unless there is a legitimate reason, notably research or IT security. A written mandate and a defined scope remain essential.
- GDPR, article 33: in the event of a personal data breach, the supervisory authority (the CNIL in France) must be notified within 72 hours.
- NIS2: the “Resilience” bill transposing the directive is under review at the French National Assembly (November 2026). It will greatly expand the number of entities subject to cybersecurity obligations.
In the event of an incident in France, the cybermalveillance.gouv.fr platform directs businesses to local service providers.
Licence: ARTEX is under AGPL-3.0. If you modify the tool and offer it as a network service, you must publish the corresponding source code. An open source licence does not, of course, authorise any illegal use.
Conclusion: the attacker no longer sleeps, and neither should your defence
ARTEX does not create new flaws. It industrialises the exploitation of the ones that already exist: reused passwords, accounts without MFA, outdated plugins, forgotten staging sites. What changes is scale and speed: three hours between first contact and compromise, hundreds of assets tested in parallel, without fatigue or pause.
The answer is not to panic, but to go back to the fundamentals with rigour:
- inventory and reduce what is exposed;
- patch actively exploited vulnerabilities first;
- enforce MFA everywhere;
- detect behaviours, not just signatures;
- automate the first response.
At Partikuls, we help our clients secure their websites and applications: exposure surface audits, maintenance and updates, WordPress hardening, MFA rollout and monitoring. Want to know what an autonomous agent would see first on your website? Contact our team for an assessment.
Sources
- jiwoochris/artex-global, GitHub
- ARTEX defence and detection guide, GitHub
- FBI says Artex AI open source hacking tool used to breach networks in under three hours, TechCrunch (16/11/2026)
- ARTEX Exposed: Infrastructure Behind the South Korea Bank Hack, Hunt.io
- ARTEX AI Pentesting Tool Used in Data Theft Attacks, The Hacker News
- Disrupting the first reported AI-orchestrated cyber espionage campaign, Anthropic
- Synthèse de la menace sur l’IA générative, ANSSI (in French)
- LLM Agents can Autonomously Exploit One-day Vulnerabilities, Fang et al. (2024)
- État des lieux de la transposition NIS2, CNB (in French)
