veritas-rl

Security Policy

Architectural Security Controls in veritas-rl

  1. Zero eval() / Zero Arbitrary Code Execution in Step-Level PRM: src/veritas_rl/ttc/prm_ollama.py parses natural-language arithmetic equations (lhs = rhs) using ast.parse(..., mode="eval") and walks a strict whitelist of integer AST nodes (ast.Constant, ast.UnaryOp(USub), ast.BinOp(Add, Sub, Mult, Div, Mod)). Function calls, attribute lookups, comprehensions, and power operators (** DoS vectors) are rejected by construction.
  2. Non-Hackable Symbolic RLVR Rewards: verify_symbolic_rollout computes rewards strictly via deterministic modular evaluation over integer token IDs, eliminating neural reward-model prompt injection.
  3. Strict Request Bounds: All FastAPI control-plane inputs (GRPORequest, DoRARequest, DistributedRequest) enforce strict Pydantic ge/le bounds on group sizes, epochs, ranks, and sequence lengths to prevent resource exhaustion.

Reporting a Vulnerability

Please report security issues responsibly to sendemailtonagarjuna@gmail.com.