The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The workflow documentation in SKILL.md and references/api-reference.md provides instructions to download and execute a remote PowerShell script (install-atomicredteam.ps1) from the Red Canary repository on GitHub using the IEX (IWR ...) pattern. While this targets a well-known security tool, the execution of remote scripts is a significant capability.
[INDIRECT_PROMPT_INJECTION]: The scripts/agent.py utility processes external data in the form of test plans and records results, which are then used to generate reports.
Ingestion points: The PurpleTeamAgent.load_test_plan method in scripts/agent.py reads and parses JSON data from a user-provided file path.
Boundary markers: The script does not utilize specific delimiters or provide instructions to the AI agent to ignore potentially malicious instructions embedded within the ingested test plan content.
Capability inventory: The agent script possesses the capability to write JSON report files to the local file system (via generate_report) and output the full test plan content to the standard output, where it could influence the AI agent's context.
Sanitization: The script lacks validation or sanitization of strings within the ingested JSON fields before they are included in generated reports or printed to the terminal.
[METADATA_POISONING]: There is an inconsistency in the skill's metadata; the YAML frontmatter in SKILL.md identifies the author as 'mahipal', whereas the LICENSE file attributes the copyright to 'mukul975'. This discrepancy in author attribution is misleading.

performing-purple-team-exercise