Quickset
A reproducible suite that compares model-file scanners on detection and false positives across a fixed corpus.
View source →Open source
Security work you can run and inspect.
Tools, benchmark fixtures, and research artifacts built from questions we wanted to answer in the open.
Model-file security / MIT
Hayward
Inspect a model file before loading it.
Hayward reads model formats without executing them. It looks for dangerous pickle behavior and suspicious content in formats such as PyTorch, GGUF, SafeTensors, ONNX, and Keras. A local CLI, CI integration, and browser demo make the same scanner useful at different points in a workflow.
pip install hayward
hayward scan ./models
Scanner benchmark / MIT
Langfail
A vulnerable application with a known answer key.
Langfail is a deliberately insecure Flask MLOps application with 82 planted vulnerabilities and 50 safe look-alikes. It gives scanners and human reviewers a realistic target with a fixed set of expected results. Run it locally for research or teaching.
More public work
Benchmarks, framework checks, and measurement.
LangChain hardening
A secure-defaults shim and a standalone OpenGrep/Semgrep rule pack for risky LangChain usage patterns.
Shim → Rules →Billhook
A public, pre-release harness for comparing prompt-injection detectors at matched false-positive rates.
View source →Prompt-guard register study
Data, code, and a paper on how prompt-injection detectors classify ordinary transactional messages.
View study →Now open source
Rowan is public.
Our source security scanner is MIT-licensed and available as an alpha. Install it with pipx install rowan-sast.