proof-checker

Pass

Audited by Gen Agent Trust Hub on Oct 27, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted LaTeX proof files and interpolates their content directly into prompts for reasoning models (e.g., Phase 1 review). It possesses write and execute capabilities that could be targeted by instructions embedded within the LaTeX data.
  • Ingestion points: Reads .tex files and proof descriptions during Phase 0 and Phase 1.
  • Boundary markers: The LaTeX content is interpolated into the reviewer prompt ([FULL PROOF CONTENT HERE]) without explicit protective delimiters or instructions to ignore embedded commands.
  • Capability inventory: The skill uses Bash, Write, Edit, and Agent tools across its workflow.
  • Sanitization: No explicit sanitization or filtering of external LaTeX content is performed before processing.
  • [COMMAND_EXECUTION]: The skill executes local system commands for document compilation and environment management.
  • Evidence: Uses pdflatex -interaction=nonstopmode to compile audit reports and shell snippets to resolve the path for research_wiki.py using git, awk, and python3. These operations are consistent with the stated purpose of mathematical document auditing.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 27, 2026, 03:07 AM