The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the agent to clone the krbrelayx repository from a public GitHub URL (https://github.com/dirkjanm/krbrelayx.git) to acquire the dnstool.py script if it is missing from the environment.
[REMOTE_CODE_EXECUTION]: Instructions direct the agent to execute Python scripts immediately after cloning them from an external repository (e.g., python3 /tmp/krbrelayx/dnstool.py), which involves running code from an unverified remote source.
[COMMAND_EXECUTION]: The skill utilizes a range of system tools such as ldapsearch, dig, nslookup, ss, and docker to facilitate network assessment and monitor attack progress. It also instructs on stopping system services and provides a handoff protocol for privileged sudo commands.
[PROMPT_INJECTION]: The skill contains instructions to override platform-level security conventions, specifically an 'OPSEC exception' that directs the agent to ignore 'Kerberos-first' requirements during coercion attacks. Additionally, the skill is susceptible to indirect prompt injection as it ingests untrusted data from tool outputs and state summaries (e.g., Responder logs, get_state_summary) and uses this data to drive subsequent actions like file writes and command construction without explicit boundary markers or sanitization.
[DATA_EXFILTRATION]: The skill is designed to capture sensitive NetNTLMv1 and NetNTLMv2 hashes and save them to the engagement/evidence/ directory, facilitating the movement and potential exposure of credential data.

auth-coercion-relay