The Agent Skills Directory

[SAFE]: The Python scripts scripts/agent.py and scripts/process.py perform static linguistic and behavioral analysis on provided text inputs. The analysis is local, uses standard libraries (re, math, collections), and lacks any network operations, dynamic execution, or persistence mechanisms.
[METADATA_POISONING]: A discrepancy was identified between the author declared in SKILL.md ("mahipal") and the copyright holder in the LICENSE file ("mukul975"). This appears to be a clerical error in the documentation rather than a deceptive attempt to obfuscate the skill's origin.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted email content through its detection engine, creating a potential attack surface for indirect prompt injection.
Ingestion points: Email text is ingested via the --file argument in agent.py and via JSON files in process.py.
Boundary markers: None are implemented to distinguish email content from system instructions.
Capability inventory: The associated scripts are restricted to file reading and printing analysis results to standard output; they cannot execute commands, access sensitive system files, or make network calls.
Sanitization: No input sanitization is performed on the ingested text, though it is only used for regex-based feature extraction and statistical scoring.

detecting-business-email-compromise-with-ai