recent-conjecture-evaluations

Warn

Audited by Gen Agent Trust Hub on Oct 26, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute a local Python script (scripts/search_inventory.py) to search for overlaps in source material and root causes.
  • Evidence found in SKILL.md and references/probe.md providing shell commands for the agent to run.
  • The references/action-policy.md also outlines a workflow for the agent to interact with GitHub by opening issues and creating draft Pull Requests, which are high-risk operations affecting external state.
  • [DATA_EXFILTRATION]: The scripts/search_inventory.py script provided with the skill allows the agent to recursively search and read the contents of files in directories provided as command-line arguments.
  • The script does not implement path validation or a sandbox to ensure that the roots arguments are restricted to the skill's output directories. If the agent is manipulated into searching sensitive locations (e.g., ~/.ssh or /etc), the script will read and output matching lines to the agent's context, potentially exposing sensitive data.
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves processing external "primary sources" (such as research papers) and "author-supplied code" (references/probe.md).
  • Ingestion points: references/probe.md step 2 ("Read enough of the primary source") and step 3 ("Label author-supplied code as SOURCE-SUPPLIED REPLAY").
  • Boundary markers: The skill suggests using labels to distinguish between source-supplied code and independent oracles, but this is a documentation practice rather than a technical enforcement mechanism.
  • Capability inventory: The agent has the ability to execute the search utility script and perform GitHub repository writes (issues/PRs).
  • Sanitization: No evidence of input validation or sanitization is present for the strings passed to the search script or the content retrieved from primary sources.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 26, 2026, 03:38 PM