Quick Take
- What: Five agentic attack frameworks (ARTEX, CyberStrikeAI, Cairn, PentAGI, HexStrike) enforce authorization only in prompts, not code
- Who is affected: Defenders hunting agent-driven intrusions and red teams assuming prompt-based scope limits will hold
- Default autonomy: CyberStrikeAI ships with approval mode off; ARTEX and CyberStrikeAI change route when blocked
- Reach: The five frameworks have drawn more than 60,000 GitHub stars combined
- What to do: Detect on behavior rather than framework strings; hunt for sensitive actions per identity in short windows
In plain English
Elastic Security Labs looked at five popular AI agent tools that security researchers and attackers use for penetration testing. These tools let a language model control scanners, shells and cloud command-line tools. Elastic found that none of the five checks permissions in the actual code, and three only tell the AI model in a text prompt not to do certain things.
What you should do: If you are a regular user, you can ignore this unless you work in cybersecurity or use these specific tools for red team work.
What happened
Mika Ayenson of Elastic Security Labs published a review of five agentic offensive frameworks on November 13, 2026. The frameworks are ARTEX, CyberStrikeAI, Cairn, PentAGI and HexStrike. The post compares how each framework handles authorization and autonomy, then pairs that with hunting queries.
Elastic notes that Google Threat Intelligence Group and Mandiant have encountered four of the five in incident response work. Microsoft attributed ARTEX use to Storm-0940.
How the frameworks work
Each framework puts a language model in charge of ordinary tools such as scanners, shells and cloud CLIs. Elastic reports that none of the five enforces authorization in code by default. Three state the authorization boundary only in the model’s prompt, where the model can reason past it.
Autonomy is the default rather than the exception. Elastic reports that ARTEX and CyberStrikeAI change route when blocked, switching proxies and User-Agent strings. CyberStrikeAI ships with approval mode off.
The five frameworks have drawn more than 60,000 GitHub stars between them. HexStrike alone exposes about 150 tool functions. Elastic reports about 390 ARTEX commits in roughly ten and a half weeks, and over 2,400 CyberStrikeAI commits in its first year.
What you should do
Elastic recommends detecting on behavior rather than on framework strings, ports or addresses, which change between releases. The hunt counts sensitive actions per identity in short windows and flags source addresses that touch at least three tool families with 20 or more requests in one minute. Baselining is necessary because build systems can trip the threshold.
Pair any address block with revoking that identity’s sessions and tokens.
What it means
Defenders hunting for agent-driven intrusion are the audience here, along with red teams that run these frameworks and assume a scope written into a prompt will hold. The research shows that prompt-based authorization boundaries are not reliable when the model itself can reason past them.



