research-wiki
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources (paper titles and abstracts via the arXiv API) and local project files (e.g.,
RESEARCH_BRIEF.md). This information is aggregated intoquery_pack.md, which serves as a context source for other skills such as/idea-creator. An attacker could theoretically embed instructions in a paper's abstract to influence future agent decisions. - Ingestion points: Paper abstracts and metadata from the arXiv Atom API; project direction and background from
RESEARCH_BRIEF.mdand other local wiki nodes. - Boundary markers: The skill uses Markdown headers to structure the aggregated data, but does not specify explicit boundary delimiters or instructions to ignore embedded commands within the
query_pack.mdoutput. - Capability inventory: The skill has access to
Bash(*),Write,Edit,WebSearch, andWebFetchtools. - Sanitization: The skill employs a
capture_filter.pyscript, though its described purpose is to filter 'operational noise' and 'transient errors' rather than specifically sanitizing against adversarial prompt injection. - [COMMAND_EXECUTION]: The skill makes extensive use of the
Bashtool to perform knowledge base operations. It executes shell commands and Python scripts to initialize, update, and query the wiki. This includes reading environment variables and local project configuration files to coordinate its behavior. - [DYNAMIC_EXECUTION]: The skill utilizes a dynamic resolution logic to locate its core Python helper (
research_wiki.py). The path to this script is computed at runtime by reading theARIS_REPOenvironment variable, checking the local.aris/installed-skills.txtfile, and querying a global configuration file in the user's home directory (~/.aris/repo). This dynamic loading from computed paths is a part of the framework's design but represents a complex execution chain.
Audit Metadata