The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to interact with and process responses from remote services, creating a vulnerability surface where untrusted data could influence agent behavior.\n
Ingestion points: Remote service responses are received through socket.recv() and pwntools.remote().recvall() as shown in references/harness-and-oracle-patterns.md.\n
Boundary markers: The skill instructions do not establish clear boundary markers or instructions to the agent to disregard command-like sequences within the received data.\n
Capability inventory: The agent has capabilities for network communication and execution of various system utilities (binwalk, jq, strings, etc.) as listed in SKILL.md.\n
Sanitization: The provided templates use regular expressions for flag extraction and basic string decoding, but do not perform sanitization to filter out potential prompt injection payloads from the service responses.\n- [DATA_EXFILTRATION]: The skill uses network operations to communicate with remote services that are not part of the standard whitelist.\n
The code templates in references/harness-and-oracle-patterns.md establish arbitrary network connections using socket.connect() and the pwntools remote function.\n- [COMMAND_EXECUTION]: The skill directs the agent to utilize several external command-line utilities for artifact analysis.\n
Tools mentioned for routine analysis include xxd, hexdump, binwalk, file, strings, and jq.

technique-ctf