bdi-mental-states
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted external RDF context and translate it into agent mental states, creating a surface for potential data-driven manipulation of agent reasoning.
- Ingestion points: The skill primarily ingests external RDF data as world state configurations (
SKILL.md) and uses Logic Augmented Generation (LAG) to generate mental states from context (references/framework-integration.md). - Boundary markers: The LAG implementation in
references/framework-integration.mdemploys structured prompt templates with clear delimiters (e.g.,## BDI Ontology,## Context to Model) to differentiate instructions from untrusted data. - Capability inventory: The skill utilizes RDF parsing, SPARQL querying, and the generation of logic-based production rules (SEMAS).
- Sanitization: The skill implements comprehensive sanitization and validation mechanisms, including the
_validate_against_ontologyfunction inreferences/framework-integration.mdand a suite of 27 SPARQL competency queries inreferences/sparql-competency.mddesigned to verify the causal and temporal integrity of the agent's mental model. - [SAFE]: The skill follows security best practices by emphasizing explainability and provenance tracking. All external URLs point to official standard bodies (W3C, FIPA, ODP), and referenced software libraries (rdflib, fipa_acl) are established packages in the semantic web and agent communities. The code snippets are instructional templates and do not execute arbitrary or dangerous commands.
Audit Metadata