formal-verification

Pass

Audited by Gen Agent Trust Hub on Oct 16, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions for downloading official toolchain managers and libraries, including the Lean toolchain manager (elan-init.sh) from elan.lean-lang.org and the TLA+ toolset (tla2tools.jar) from its official GitHub repository.
  • [COMMAND_EXECUTION]: The skill executes verification tools and compilers such as lake (Lean build system), java (running TLA+ SANY and TLC checkers), and bend (proof checker) to validate formal specifications and models.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and audit AI-generated proofs and models, creating an attack surface for indirect prompt injection if the inputs contain malicious instructions.
  • Ingestion points: Reads external proof and specification files such as .lean, .tla, and .bend files (e.g., evals/fixtures/lean-agent-proof/Dedup.lean).
  • Boundary markers: Instructions emphasize freezing human-approved statements to prevent the agent from weakening properties during the proof process.
  • Capability inventory: Includes execution of build systems (lake), model checkers (TLC), and custom checkers (bend) which may involve subprocess calls and file system operations.
  • Sanitization: The skill implements a robust mitigation strategy by providing specific 'Audit' and 'Acceptance gate' procedures (e.g., lake build --wfail, #print axioms, and grepping for escape hatches like sorry, axiom, or @unsafe) to ensure no unverified assumptions or instructions have bypassed the checker.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 16, 2026, 05:06 PM