The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's Quick Start section directs users to run npx agentic-jujutsu. This command fetches and executes code from the public NPM registry. The package is not associated with a recognized trusted organization or the provided vendor infrastructure.
[DYNAMIC_EXECUTION]: Multiple implementation examples, such as the SelfImprovingAgent class and the adaptiveDeployment function, demonstrate a pattern where the skill retrieves recommendedOperations from an AI-powered suggestion engine (jj.getSuggestion) and executes them in a loop. This behavior allows dynamically generated strings to control shell execution without clear constraints.
[INDIRECT_PROMPT_INJECTION]: The skill contains an attack surface for indirect prompt injection where untrusted data can influence agent actions.
Ingestion points: Untrusted data enters the context through jj.getSuggestion(taskDescription) and jj.finalizeTrajectory(score, critique) in SKILL.md.
Boundary markers: There are no explicit delimiters or instructions provided in the documentation to prevent the AI from following malicious instructions embedded in the task descriptions or critiques.
Capability inventory: The skill uses jj.execute() to run shell commands, providing a high-impact capability that can be manipulated by malicious input.
Sanitization: The provided code snippets do not show any validation, sanitization, or human-in-the-loop verification of the operations before they are executed.

agentic-jujutsu