AI/ML●●Solid
AST-guard A gradient-immune structural guard against RL reward hacking
Gradient-immune AST analysis that RL models can't optimize against through backpropagation.
Big BrainNiche Gem
thinking-nick
3025d 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 – Fast, zero-cost structural checks for LLM code execution
by thinking-nick·Jun 8, 2026·2 points·0 comments
AI Analysis
●●●BangerBig BrainWizardry
Deterministic AST analysis catches AI code bypasses that LLM reviewers miss, verified on 77k+ samples.
Strengths
- •AST-based deterministic checks can't be talked into compliance like LLM reviewers
- •External dataset validation (MALT, School of Reward Hacks) shows real rigor
- •Sub-10ms zero-cost gating enables production deployment without latency tax
Weaknesses
- •Detection rates vary widely (34.5-95%) depending on attack category
- •Python 3.11+ requirement limits deployment in legacy environments
Category
Target Audience
AI safety engineers, teams running LLM-generated code in production
Similar To
TRACE · EvilGenie · RewardHackWatch
Similar Projects
Developer Tools●●●Banger
Pre-execution verification for LLM-generated agentic workflows
Type-safe AST verification for AI workflows before they corrupt your CRM or delete production data.
Big BrainSolve My Problem
jaredwaxman
454mo ago
Security●●Solid
ÆTHERYA Core – deterministic policy engine for governing LLM actions
Fail-closed guardrails for LLM actions with cryptographic approval and audit chains.
Big BrainNiche GemSolve My Problem
RobertMihai
104mo ago
Security●●Solid
Access-aware text-to-SQL – stop LLM agents overfetching data
Deterministic guard prevents overfetching when 121 adversarial tests pass.
Big BrainNiche Gem
dimitarst
1117d ago
AI/ML●●Solid
Nexa-Gauge – LLM eval framework, now with self-hosted model support
Cache-aware LLM eval with self-hosted model support beats Ragas on flexibility.
Solve My ProblemSlick
Sardhendu
202mo ago
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
313mo ago