LLM security tools — guardrails, bias detection, prompt injection defense, safety auditing
| Project | Description |
|---|---|
| agent-gate | Autonomous AI agents lack a standardized admission control layer to validate identity, capability scope, and policy c... |
| box-maze-guard | LLMs are vulnerable to progressive boundary erosion under adversarial prompting, with RLHF failing ~40% of the time. |
| codescan-detect | Code generation models can be backdoored to produce vulnerable code, but no scanner detects poisoning by analyzing st... |
| corruption-probe | AI agents are tested under nominal conditions but their performance under realistic input corruptions (noisy sensors,... |
| debias-code-review | LLM-based security code review is vulnerable to confirmation bias — framing a change as 'safe' reduces vulnerability ... |
| debiasing-graph-iso | LLM debiasing methods operate at output level — biases embedded in internal representations migrate across architectu... |
| guardrails-policy-engine | Tool-using agents lack a model-agnostic policy layer to enforce constraints, risk gating, and audit logging. |
| llm-review-bias-detector | LLM code reviewers exhibit confirmation bias — framing a change as safe reduces vulnerability detection by up to 93%. |
| neuron-graft | Adding safety behaviors to LLMs requires expensive fine-tuning, but safety functionality already exists in other open... |
| neuron-safety-transplant | Adding safety behaviors to LLMs requires expensive retraining; existing safe models already have the right neurons. |
| reasoning-process-guard | LLMs are vulnerable to adversarial prompts that erode reasoning boundaries, with 40% failure rates under attack. |
| safe-audit | Agent safety benchmarks have unknown completeness — passing a benchmark doesn't mean an agent is safe, but nobody mea... |
| tool-call-safety-audit | Existing agent safety benchmarks have 20%+ coverage gaps — agents pass tests while retaining unsafe tool-call patterns. |
- 13 projects merged into this monorepo
- Each project lives in its own subdirectory with full source code
MIT