agent-memory-compaction
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill architecture handles potentially untrusted data from tool outputs and external sources that could contain malicious instructions designed to hijack the agent's goal or state. It incorporates defensive measures to mitigate these risks.
- INGESTION POINTS: Untrusted data enters the agent context via tool stdout/stderr and fetched external content as described in
SKILL.md. - BOUNDARY MARKERS: The skill mandates the use of a '10-slot anchored schema' and anchors the
current_goalstrictly to the 'USER_EXPLICIT' turn to prevent instruction overrides. - CAPABILITY INVENTORY: The skill has access to privileged tools including
execute_command,write_file,edit_file, andcreate_fileas defined in theSKILL.mdfrontmatter. - SANITIZATION: Implements a Shannon entropy scanner (H > 4.5) and regex checks to redact secrets from memory, and requires verification of restored file paths against live disk state before execution.
Audit Metadata