The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external binaries and interact with remote network services, which could contain malicious instructions designed to influence the agent.
Ingestion points: Handout binaries (chall), library files (libc.so.6), and remote service network streams (nc host port) as described in the triage and pwntools loop sections of SKILL.md.
Boundary markers: The instructions do not include delimiters or specific "ignore" instructions for content ingested from the target binaries or remote services.
Capability inventory: The agent is granted capabilities to execute binaries locally (process), connect to remote hosts (remote), modify binary metadata (patchelf), and manage containers (docker) as specified in the tools and Method sections.
Sanitization: There is no evidence of sanitization or validation of data received from the remote service or extracted from the binary before it is processed by the agent.
[DYNAMIC_EXECUTION]: The skill provides Python code templates that are used to generate and execute exploit scripts at runtime.
Evidence: SKILL.md contains a Python template utilizing the pwntools library to instantiate process() and gdb.debug() objects for local binary execution.
[COMMAND_EXECUTION]: The skill instructs the agent to execute powerful command-line tools to modify system state or binary properties.
Evidence: SKILL.md provides explicit commands for using patchelf to set interpreters and rpaths, and docker commands to build and run containers with elevated privileges (--cap-add=SYS_PTRACE).

ctf-pwn