The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted conversation history to generate structured summaries and evaluation probes, creating a surface for indirect prompt injection attacks.\n
Ingestion points: The scripts/compression_evaluator.py script ingests raw conversation history through the ProbeGenerator class and truncated content spans via the StructuredSummarizer.update_from_span method.\n
Boundary markers: The implementation lacks explicit delimiters or specific instructions to the agent to ignore potentially malicious instructions embedded within the processed history.\n
Capability inventory: The evaluation script is focused on text processing and heuristic scoring; it does not contain capabilities for network requests, file system modifications, or shell command execution.\n
Sanitization: The script extracts information using regular expressions but does not perform sanitization, escaping, or validation on the resulting content before it is used in subsequent contexts.

context-compression