The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The scripts scripts/agent.py and scripts/process.py ingest and parse SAML federation metadata from remote URLs. This creates a surface for indirect prompt injection if the metadata source is compromised, potentially allowing malicious content to enter the agent context.
Ingestion points: The functions fetch_federation_metadata in scripts/agent.py and validate_adfs_metadata in scripts/process.py fetch external XML metadata.
Boundary markers: The skill does not implement explicit delimiters or warnings to instruct the agent to ignore instructions potentially embedded within the XML fields.
Capability inventory: The skill possesses network communication capabilities (requests) and file writing capabilities (json.dump), which are used for generating configuration reports.
Sanitization: Input validation is limited to standard XML parsing; the skill does not perform deeper sanitization of the values extracted from the SAML assertions before processing.
[METADATA_POISONING]: There is a discrepancy between the author metadata field in SKILL.md ('mahipal') and the copyright attribution in the LICENSE file ('mukul975'). While this appears to be a clerical or template-related inconsistency rather than a malicious attempt to deceive, it represents a metadata quality issue for auditing purposes.

building-identity-federation-with-saml-azure-ad