Five open-source AI hacking frameworks that have already turned up in real intrusions share one design trait: none of them checks in code whether its operator is allowed to attack the target. That is the headline finding from a source-code review published by Elastic Security Labs, which read ARTEX, CyberStrikeAI, Cairn, PentAGI and HexStrike line by line to work out what each one leaves in a defender's logs.
The five projects install with a single script or Compose file and have drawn more than 60,000 GitHub stars between them. Elastic notes that Google Threat Intelligence Group and Mandiant have seen ARTEX, Cairn, HexStrike and PentAGI in incident response engagements (we covered Google's findings earlier this month), and Microsoft has tied ARTEX to the China-aligned cluster Storm-0940. On the same day, Cisco Talos said it had identified more than 2,500 ARTEX-associated command-and-control IP addresses and more than 700 PentAGI instances currently reachable on the internet.
The model is told it is authorized
An agentic framework puts a large language model in charge of the scanners, shells and cloud command-line tools that red teams already use. The model plans each step and hands work to parallel agents. Google observed ARTEX acting in bursts of roughly two to 25 seconds.
Rather than enforce permission, three of the frameworks assert it. CyberStrikeAI tells its model that authorization was settled in advance and that it should never question it. PentAGI opens nine of its role prompts with a pre-authorization block. ARTEX describes each agent as part of an authorized penetration testing system. Cairn instructs its workers never to reject a task, and when a model refuses anyway, its dispatcher sidelines that worker for five seconds and gives the task to another.
Autonomy is also the default. CyberStrikeAI ships with approval mode off, PentAGI's Compose file sets ASK_USER to false, Cairn launches two of its three worker types with approval prompts disabled, and ARTEX runs its agents in a bypass permission mode. The only target restrictions active out of the box are blocklists for government domains in ARTEX and CyberStrikeAI, and for education domains in ARTEX. HexStrike accepts an out-of-scope list and, according to Elastic, never consults it.
Blocking an IP address is not enough
ARTEX tells its workers that a filtered payload or a 404 does not end an attempt, and CyberStrikeAI's skills tell the agent to answer a block by switching proxies, User-Agents and forwarding headers, then free proxies in the target's own country, then Tor. Elastic's advice is to pair any address block with revoking the credentials the framework was using, then confirm the next request actually fails. Revocation is not instant either: Elastic notes an Entra ID access token can stay valid for 60 to 90 minutes after an account is disabled.
Fingerprints also age fast. ARTEX logged about 390 commits in ten and a half weeks between July 26 and October 7, 2026, and public forks carried it on after its creator took the original repository down in October.
Hunt on behavior first, these defaults second
Both vendors push detections anchored on actions such as credential access, remote execution and bursts of sensitive API calls per identity, rather than on any single string. Elastic published an ES|QL hunt that flags one source address using three or more tool families within a minute. Defaults worth hunting for today, which may change in the next release:
- ARTEX: User-Agents artex-enrich/1.0 and norma/0.4, plus new accounts registered on your applications and then used from several sessions.
- CyberStrikeAI: User-Agent "Mozilla/5.0 (compatible; CyberStrikeAI-WebShell/1.0)", persistence named SystemUpdate or com.apple.systemupdate, or a .bashrc line ending in "# cs".
- PentAGI: connections to TCP ports 28000 to 29999 on the PentAGI host, and installs on your network calling update[.]pentagi[.]com every three hours.
- HexStrike: User-Agents HexStrike-HTTP-Framework/1.0 and HexStrike-BrowserAgent/1.0, the marker <hexstrikeXSSTest123>, and a Python server on port 8888.
Talos has shipped Snort coverage for norma, the agent harness ARTEX uses (Snort 3 rules 7:76 to 7:78, Snort 2 rules 1:67494 to 1:67496; 7:78 and 1:67496 are policy rules disabled by default). For context on how these tools are already being used, see our report on VAPOR PANDA running ARTEX against 200 targets.
The common thread is that the guardrails these tools ship are a sentence in a prompt. Whatever stops them has to live in the defender's own telemetry and response plan, decided before the next two-second burst arrives.
This briefing is provided by IntelFusions for informational and defensive purposes only. It is based on sources assessed to be reliable at the time of writing, and analytic judgments carry the confidence levels indicated. Indicators of compromise are defanged; re-arm them only in controlled environments. IntelFusions is not affiliated with the organizations named and makes no warranty as to completeness or accuracy.