task-observer

Warn

Audited by Gen Agent Trust Hub on Aug 27, 2026

Risk Level: MEDIUMPERSISTENCECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
  • [PERSISTENCE]: The skill implements a mechanism for setting up autonomous background tasks to perform periodic audits and updates.
  • Evidence: The references/weekly-review.md file contains instructions for the agent to register a weekly-skill-review task using system schedulers such as cron, create-shortcut, or set_scheduled_task.
  • [COMMAND_EXECUTION]: The skill makes extensive use of shell utilities to manage its internal logging system and handle file updates across different environments.
  • Evidence: SKILL.md and references/weekly-review.md utilize commands including grep, sed, awk, chmod, cp, and diff to number observations, reformat log files, and adjust file permissions for staged updates.
  • [INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and processing unstructured session data, user feedback, and deliverables to generate rules for other skills, creating a potential vector for indirect injection.
  • Ingestion points: The skill reads session history, user-provided handoff documents, and the skill-observations/log.md file as described in SKILL.md and references/environments.md.
  • Boundary markers: The skill uses a structured Markdown format for logs but does not enforce technical escaping or delimiting of the content processed from external inputs.
  • Capability inventory: The skill is capable of writing new skill files, modifying existing instructions, and executing shell commands through the agent.
  • Sanitization: There are instructions for the agent to manually 'strip specifics' from observations in references/skill-authoring.md, but no automated sanitization is present.
  • [METADATA_POISONING]: The skill's metadata includes imperative directives designed to override the agent's standard skill invocation logic and modify the local environment's configuration.
  • Evidence: The description field in SKILL.md contains an 'IMPORTANT' instruction requiring the skill to be invoked at the start of every session and directs the agent to add activation instructions to the project's CLAUDE.md file.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Aug 27, 2026, 02:33 PM