opendeviation-eval-metrics
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEPROMPT_INJECTIONPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The
SKILL.mdfile includes a 'Self-Evolving Skill' directive that explicitly instructs the agent to 'fix this file immediately' if instructions are incorrect or parameters have drifted. This allows for the modification of system-level instructions based on agent experience or potentially malicious inputs. - [PERSISTENCE]: By directing the agent to modify its own
SKILL.mdfile, the skill enables a form of persistence where changes to the agent's behavioral logic can be made and maintained across different sessions. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data in
.npyand.jsonlformats. This creates a vulnerability surface where maliciously crafted data could attempt to influence the agent's behavior or trigger the self-modification logic. - Ingestion points:
scripts/compute_metrics.pyprocesses user-provided binary and structured data files. - Boundary markers: No specific delimiters or 'ignore' instructions are used when interpreting the data content.
- Capability inventory: The skill utilizes
Bash,Read,Grep, andGlobtools, providing a significant capability set if the agent is subverted. - Sanitization: The data is processed using standard numerical libraries; no sanitization exists for natural language instructions that might be embedded in data or metadata.
Audit Metadata