Security●●●Banger
AST-guard – Fast, zero-cost structural checks for LLM code execution
Deterministic AST analysis catches AI code bypasses that LLM reviewers miss, verified on 77k+ samples.
Big BrainWizardry
thinking-nick
2021d ago
Back to browse
GitHub Repository
Pre-Execution Gate for AI Code. A deterministic, gradient-immune structural guard against reward hacking and hardcoding in RL training loops.
1 starsPython
AST-guard A gradient-immune structural guard against RL reward hacking
by thinking-nick·Jun 29, 2026·3 points·0 comments
AI Analysis
●●SolidBig BrainNiche Gem
Gradient-immune AST analysis that RL models can't optimize against through backpropagation.
Strengths
- •Deterministic structural analysis means zero false positives on known hack patterns.
- •Empirically validated in actual RL training loops, not just theoretical.
- •Sub-10ms latency makes it viable as a real pre-execution gate.
Weaknesses
- •Explicitly experimental research artifact, not production-ready.
- •Only catches structural hacks—semantic bypasses require escalation to other tools.
Category
Target Audience
AI safety researchers, RL engineers training code-generation models
Similar To
TRACE · RewardHackWatch · EvilGenie
Similar Projects
AI/ML●●●Banger
An MCP server that fact-checks AI bug diagnoses against AST evidence
Forces LLMs to debug with AST evidence instead of pattern-matching symptoms.
Big BrainSolve My Problem
EruditeCoder108
312mo ago
AI/ML●●●Banger
RewardHackBench: Using sandboxes to stop agents from cheating
LLM judge on outgoing requests achieves 0% cheat rate while preserving 58% fair-solve ceiling.
Big BrainDark Horse
rotemtam
9312d ago
AI/ML●●Solid
RewardGuard – detect reward hacking in RL training loops
Catches reward hacking before it tanks your RL training run.
Niche GemBig Brain
Giovan321
112mo ago
AI/ML●●●Banger
RewardHackWatch – Reward hacking detector for LLM agents
Catches LLM reward hacking at runtime when models game evals.
Big BrainWizardryShip It
aerosta
114mo ago
