Report vulnerabilities responsibly to the repository owner by email to g@abejar.net -- do not open public issues.

This project is a defensive adversary-emulation harness intended for blue-team training and detection-engineering. Its techniques are stubs that emit synthetic evidence into a sandboxed workdir; they do not perform real attacker actions. Specifically:

• No technique opens network sockets to attacker-controlled infrastructure.
• No technique modifies the real Windows registry, schedules real tasks, dumps real LSASS memory, or extracts real secrets.
• No technique writes outside the configured workdir; all paths flow through Technique._safe_join.
• A deny-list (DENY_LIST) rejects technique IDs associated with destructive impact (data destruction, ransomware, inhibit recovery, defacement, firmware corruption, disk wipe). Both the registry and the executor enforce this.
• The LLM planner is constrained to the registry's published IDs and is filtered against the deny-list before execution.

• The harness assumes operation on a developer or analyst workstation, possibly inside a CI sandbox.
• It does not protect against an operator who is determined to extend the framework to perform real-world attacker actions; that is outside the project's defensive scope. Removing the deny-list, re-implementing a technique stub to actually execute, or wiring the framework to a real C2 server is not supported and would require forking.
• The LLM advisor receives a description of the registry (technique IDs, names, descriptions) and the user-supplied objective. No personal data, evidence, or system state is transmitted.

If you discover this project being used to coordinate real attacker activity, contact the repository owner. The project's name and README make its defensive purpose unambiguous.