From Report to Tested Detection in Hours: The ARTEX Signature Pack

The researcher MalwareMustDie released an open ARTEX detection pack in seven formats, built on behavior and a repeatable playbook, because attackers now turn weeks into hours.
The researcher MalwareMustDie released a public detection pack for ARTEX, the open-source AI pentest framework that criminals now run against real targets. He published it quickly because campaigns built on this framework are compressing intrusion timelines from weeks into hours.

The pack comes in seven formats: Sigma, including five behavioral correlation rules, Sentinel KQL, Suricata 7, Snort 3, osquery, YARA and a STIX 2.1 bundle. The IOCs expire after a year by design, and the behavioral rules don’t expire. If you read this as an IOC list, you’ve missed the point.
“The real release is the workflow: how to turn the next ARTEX-style attack into detections as fast as the attacker moves.” reads the report. “ARTEX is an LLM harness. Swap the model, prompts or infra and every IOC changes. What changes much less is how an agent behaves: machine-speed recon chains, ~1/s Password Sprays with rotating UAs, self-correcting API retries, systematic record enumeration. So the pack is behavior-first.”
ARTEX is a framework that connects AI models with tools. If attackers change the AI model, prompts or servers, the signs of an attack can also change. Many ARTEX indicators found on compromised systems are created by the AI model or the tools it uses, rather than being built directly into the framework.
For this reason, the researcher focused on sequences of suspicious actions rather than individual strings. These include automated reconnaissance at high speed, password attacks with about one attempt per second and changing user agents, API requests that automatically retry with corrections, and systematic searches through database records.
Every indicator in our detection plan goes through eight checks. These cover its source, reliability, classification, format and testing. Confirmed ARTEX indicators trigger alerts first. More generic indicators come with a warning and are never used alone to raise an alert. For example, the use of a curl user agent is common in legitimate traffic, so it should not trigger a security alert by itself.
Every indicator in the playbook passes eight gates, from source check to tier, format and testing. Confirmed ARTEX indicators alert first, and generic strings carry a warning and never fire alone. A curl user agent appears in plenty of harmless traffic, so nobody should get paged for it.
“Act on what the identity does, not where it comes from.” continues the report. “The Global Administrator grant came from Microsoft-internal private IPv6, so the ConnectWise app-ID rule now fires on role self-assignment or credential addition regardless of source IP. Source IP is enrichment only.”
MalwareMustDie tested the playbook on four campaigns, with an LLM helping and humans signing off every verification step. The ConnectWise Cloud Backup case from October 2026 shows why. A vendor’s app credential was compromised, and service principals then gave themselves directory roles, with one becoming Global Administrator from a Microsoft-internal private address. Microsoft said its own systems weren’t compromised, and attackers planted secrets that expire in 2028. A rule that trusts the source address would miss this.
The Korean financial case taught a different lesson. Attackers pulled about 25,000 records over roughly 30 hours, around 14 a minute, from peripheral systems like a broker loan inquiry service and an employee app. Volume thresholds miss that, so we count distinct records read per identity over time. The Japanese wave, reported by JPCERT/CC and GMO, exploited known bugs and misconfigurations, with no zero-days.
The expert also left things out on purpose. No backdoor trigger values, no private individuals and no victim names appear in the pack. Indicators he couldn’t verify wait in a hold queue until someone confirms them.
For network teams, the researcher added an experimental Zeek approach. One layer checks who you’re talking to, and another tracks what’s said and how often. Zeek can’t see sign-ins to Entra ID, though, because that traffic never crosses your perimeter.
The pack has moved since release. He updated it to v1.1.3 after traffic tests, added data from the FBI flash report that came out the same day, and confirmed Sophos X-Ops indicators that had been on hold. A chronology of public ARTEX reporting from October and November is in Appendix F. The quick Shodan scan on release day wasn’t comforting.
The full report includes the detection playbook, rules and indicators of compromise (IoCs). The expert encourages other researchers to test them, identify weaknesses and report both false positives and true positives.
Attackers use LLMs to complete in hours what once took weeks. Defenders must keep up with this speed while ensuring that humans remain involved in the process.
Defenders can use LLMs to compress “new report → tested detection” into hours too, as long as humans keep the verification gates.
Follow me on Twitter: @securityaffairs and Facebook and Mastodon
Pierluigi Paganini
(SecurityAffairs – hacking, ARTEX)

