The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines workflows for intelligence gathering and vulnerability analysis that require ingesting data from external, potentially untrusted targets. This creates an attack surface where an adversary could embed malicious instructions in service banners, web content, or metadata to attempt to influence the agent's behavior.
Ingestion points: Findings from intelligence gathering tools (Nmap, service enumeration) and target-context parameters provided by the user in SKILL.md.
Boundary markers: While the skill emphasizes authorization and audit trails, it does not explicitly instruct the agent to use delimiters or ignore instructions embedded within data retrieved from targets.
Capability inventory: The skill integration map includes powerful capabilities such as bash tool execution (Nmap, Nuclei, sqlmap, ffuf) and the generation of exploitation scripts.
Sanitization: No specific instructions are provided for the agent to sanitize or validate content retrieved from external targets before it is used in subsequent reasoning or tool execution.
[DYNAMIC_EXECUTION]: The skill provides instructional templates for generating Python scripts that interact with and execute local binary files for binary exploitation tasks.
Evidence: The pwntools template in SKILL.md demonstrates how to load an ELF binary and spawn a local process for testing using ELF('./challenge') and process('./challenge').

offensive-security