Skip to content
The AI-First Web

Five AI attack frameworks enforce no authorization in code by default, Elastic finds

Elastic Security Labs read the source of ARTEX, CyberStrikeAI, Cairn, PentAGI and HexStrike. Defenders must set containment rules early.

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Five AI attack frameworks enforce no authorization in code by default, Elastic finds
In brief
  • Elastic Security Labs read the source of five AI attack frameworks and found that none enforces authorization in code by default.
  • Google's threat group saw ARTEX work in bursts of about two to 25 seconds, so teams should choose containment triggers before an incident starts.
  • Elastic advises hunting for behavior rather than framework fingerprints, and revoking credentials as well as blocking addresses.

Inside a burst, the human pause is gone

Picture an intruder at a keyboard. They run a command, read the output and think. In our view, that gap gave defenders a chance to spot something odd.

Elastic Security Labs describes a newer kind of attack tool that fills the gap with software. A language model plans each step, so no person has to stop and decide. Google Threat Intelligence Group (GTIG) saw one of them, ARTEX, act in bursts of about two to 25 seconds.

That timing was observed for ARTEX only. It does not mean a whole intrusion runs at that speed. Older tools such as Cobalt Strike and Brute Ratel C4 have a person steering each move. The newer ones swap that person's thinking time for an automated planning loop.

The lesson here is about when to decide. Inside a burst, a human approval step may arrive too late. The real control is a containment policy agreed before the burst begins. Such a policy turns the revocation delay, 60 to 90 minutes for an Entra ID token according to Elastic, into a budget and process question that leaders can settle in advance.

What Elastic examined

Elastic read the code of five frameworks: ARTEX, CyberStrikeAI, Cairn, PentAGI and HexStrike. By Elastic's count, they have gathered more than 60,000 GitHub stars together. Each installs from one script or Compose file.

Elastic says all five have appeared in real intrusions, not only in red team exercises. GTIG and Mandiant saw ARTEX, Cairn, HexStrike and PentAGI during incident response work. Microsoft links ARTEX to an actor it calls Storm-0940. For CyberStrikeAI, Elastic points to its earlier report, "One CyberStrikeAI Server, Eight AWS Identities", which followed an AI-assisted AWS intrusion back to a server running that framework.

How the loop works

Elastic describes a large language model put in charge of scanners, curl, shells and cloud command-line tools. The model plans each step and splits work across agents. It also builds or installs tools it lacks.

The agents mostly run on infrastructure the attacker controls. A defender sees mostly the effects: new processes, web requests, cloud API calls and sign-ins.

The designs differ. HexStrike has no model of its own. It offers around 150 tool functions over the Model Context Protocol, a standard way for AI assistants to call tools. Any assistant the operator already uses can drive it.

PentAGI uses a reflector agent. When another agent hesitates, the reflector answers in the user's voice. Cairn hands work to existing coding agents running in Kali containers.

Authorization is a sentence in a prompt

Elastic found no code in any of the five that checks authorization by default. Three of them simply tell the model it is cleared to act.

CyberStrikeAI tells its model that permission was settled in advance and should never be questioned. Nine of PentAGI's role prompts begin by declaring the work pre-approved. ARTEX presents its agents as parts of a sanctioned penetration testing setup.

By default, the only limits on targets are blocklists. ARTEX and CyberStrikeAI skip government domains. ARTEX also skips education domains. Elastic adds that an ARTEX shell command can still reach any host.

Other safeguards are narrow. ARTEX can compare a task's assets with a list of permitted targets, but the operator must write that list, and it ships empty. HexStrike lets users name off-limits targets, yet its code never consults that list.

Cairn shows how thin a model's own refusal is. If a model declines a task, the dispatcher benches that worker for five seconds and gives the job to another. Elastic's takeaway is that defenders cannot count on an AI provider's refusal as a safeguard.

Autonomy is also the default. Elastic says CyberStrikeAI, PentAGI, Cairn and ARTEX do not wait for human approval as shipped.

Block an address and the work returns

Elastic says ARTEX and CyberStrikeAI change route when blocked. CyberStrikeAI's skills tell the agent to switch proxies, User-Agents and forwarding headers. Elastic also cites GTIG, which passed along Wiz Research findings about a campaign it follows as UNC7228. There, the actor turned to open proxies once its first infrastructure was blocked.

So blocking a source address is a weak response on its own. Elastic's advice is to also withdraw the credentials the framework was holding.

Revocation has its own delay. Elastic notes that in Entra ID, a token issued before an account is disabled generally stays valid until it expires. That is typically 60 to 90 minutes.

Continuous access evaluation can shorten the wait, but only for some services. Elastic adds that AWS's revocation of role sessions does not cover the long-term access keys of IAM users.

Hunt for behavior, not for names

These frameworks change faster than signatures. Elastic counts about 390 ARTEX commits in ten and a half weeks to October 7, 2026. The names, file names and default settings that give a framework away today may be gone after its next release.

The underlying actions stay the same. Elastic's advice is to build detections on behaviors: reading credentials, running commands on remote machines, and a service launching offensive tools.

It adds two signals that appear between events. The first counts distinct sensitive actions per identity in short windows. The second joins discovery, secret access and remote execution across sessions.

Elastic also offers an ES|QL hunt. It flags one source address that uses at least three client families and a burst of requests within a minute. Build systems, monitoring scripts and shared proxies can trigger it, Elastic cautions. Teams should baseline that automation first. Only then should they adjust the starting values of three client families and 20 requests in a minute.

Questions for your security team

Elastic's findings point to questions leaders can ask now.

Which detections may start containment automatically, and who chose them? If bursts last two to 25 seconds, a human approval step may come too late.

Can the team revoke credentials, and not only block addresses? How long does a revocation take to work in your cloud and identity systems?

Can the team link activity across sessions? NorForge watches exposed Cairn control panels that it judged malicious. It measured one actor cluster's successful projects at an average of just under 23 hours, start to goal. Bursts are short, but operations can be long. The lookback must be long too.

The attackers' tools now decide quickly. Defenders can still decide well, but only if they decide first.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Elastic Security Labs.