ARTEX and AI-driven cloud intrusions

November 18, 2026Zainab Ali Majid & Katarina Zotovic · DFIR Investigations

The past weeks have been some of the busiest we’ve seen in the industry. Much of that is down to ARTEX, a public AI penetration-testing framework now being used in cloud compromises at scale.

In one recent ARTEX case we investigated, an intrusion that would normally take days or weeks took minutes. By the time an analyst would typically have triaged the first alert, the attacker had accessed thousands of secrets, session tokens, API keys, and credentials.

The attacker left a recognizable signature: file paths referencing ARTEX-main, resources labeled with ARTEX intent IDs such as i1234, and requests carrying the norma/0.4 user agent. Those requests arrived at machine speed, with a median gap of just 0.25 seconds.

Across these cases, the pattern is consistent: broader access in far less time, more flexible lateral movement compared to traditional reconnaissance, and a forensic trail that is harder to reconstruct.

Big cloud targets

The victim was a web application running on a large AWS estate, with production behind a public sign-up page. The attacker created an account like any trial user would, then moved through the application layer down to the underlying AWS infrastructure.

The initial reconnaissance was typical: confirm connectivity, identify the version and current role, check if the role can escalate privileges, access further resources, and recursively repeat this process. Human attackers have done the same, but on significantly longer timescales.

Messy execution

An AI-driven operator can explore numerous paths at once, keeping what works and leaving behind evidence from those which fail. In this matter, we tracked nearly a dozen separate attack branches with varying levels of success and failure:

  • Succeeded: access to the production application, staging platform, internal admin application and several AWS accounts.
  • Failed: attempts to access EKS clusters and move laterally into additional cloud environments.

Responders have to understand every branch and confirm that the failed ones really failed. That slows analysis, makes it harder to pin down which data is at risk, and widens the scope of containment.

Actions from limited infrastructure

One detail surprised us during the investigation: most of the intrusion came from a single IP address, active on that one day and nowhere else in the data.

Human-only cloud intrusions usually involve infrastructure sprawl: rotating VPNs, residential proxies and sessions hopping between providers over days. Defenders have learned to scope attacks by following those connections.

Our working theory is that speed reduces the need for that sprawl. With little dwell time to conceal, there may be less reason to rotate infrastructure. In this case, the sprawl was in the attacker’s actions. We found hundreds of attacker accounts, API keys, OAuth and MCP clients, and other attacker-created artifacts, all from the same source IP.

Hard to track

To scope this investigation, we searched hundreds of millions of log records from over 20 sources, including web requests, application level audit logging, CloudTrails across AWS accounts, EKS, VPC flow logs, and more. Three things made it hard.

First, the attack spanned both the application and the infrastructure. AWS evidence alone wasn’t enough; we had to collect and parse every available application log.

Second, ARTEX searches recursively. It branches, fails, adapts and branches again, so a single session fans out into many separate threads of activity.

Third, every failed branch has to be closed out. Proving what an attacker didn’t do matters as much as knowing what they did. When an attack spreads this wide, gaps in logging make that hard to establish.

Attacks are getting faster. Investigators have to keep up.