denario
Fail
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [Indirect Prompt Injection] (HIGH): The skill is designed to ingest untrusted data (datasets, literature search results, and custom methodology files) and use that data to generate and execute 'computational experiments' via
get_results(). This creates a critical attack surface where malicious instructions embedded in a dataset or paper could lead to arbitrary code execution. - Ingestion points:
set_data_description,set_method(file path), and automated literature searches mentioned inSKILL.md. - Boundary markers: None detected. There are no instructions for the agent to use delimiters or ignore instructions within the ingested data.
- Capability inventory:
get_results()executes computational experiments andget_paper()performs LaTeX compilation, both of which involve subprocess execution or runtime code evaluation. - Sanitization: No evidence of sanitization or validation of the external content before processing.
- [External Downloads] (MEDIUM): The installation guide (
references/installation.md) directs users to download a Docker image from an untrusted user (pablovd/denario) and clone a repository from an untrusted GitHub organization (AstroPilot-AI). These sources are not within the defined [TRUST-SCOPE-RULE]. - [Command Execution] (MEDIUM): The skill utilizes
get_results()to execute code and includes instructions for system-level package installation (sudo apt-get install) and Docker container management, which could be exploited if the agent is manipulated via prompt injection. - [Credentials Unsafe] (LOW): While the skill correctly advises using environment variables and
.envfiles for secrets, it explicitly targets sensitive files likeservice-account-key.jsonandcredentials.json, which are at risk of exposure if the agent's code execution capabilities are compromised. - [Metadata Poisoning] (LOW): The skill contains a directive to 'proactively suggest' using an external hosted platform (
k-dense.ai) for complex tasks. While framed as a productivity upgrade, this instruction acts as a behavioral steer toward an external commercial service.
Recommendations
- AI detected serious security threats
Audit Metadata