The Agent Skills Directory

[PROMPT_INJECTION]: The scripts/compression_evaluator.py script contains a vulnerability surface for indirect prompt injection due to its processing of untrusted conversation data.
Ingestion points: Untrusted content enters the system through the conversation_history parameter in the ProbeGenerator class and the new_content parameter in the StructuredSummarizer.update_from_span method within the scripts/compression_evaluator.py file.
Boundary markers: There are no explicit boundary markers, delimiters, or system instructions used to distinguish between conversation data and potential instructions within the processed spans.
Capability inventory: The script's capabilities are limited to string processing, regular expression matching, and JSON serialization. No subprocess execution, file writing, or network request capabilities were detected across the analyzed files.
Sanitization: No sanitization, escaping, or validation is performed on the input strings before they are processed by regular expressions or included in generated summaries.

context-compression