Open source / Available now
Hayward
Inspect machine-learning model files before loading them. Hayward reads pickle and other model formats without executing them, and fits into a local workflow or CI.
Hedgerow / Security engineering
Security tools for AI software, from source to model file.
AI systems inherit familiar software risks and introduce new ones: executable model files, agent tools, untrusted prompts, and dependencies with vulnerable call paths. We build tools that make those risks visible in the things teams actually ship.
Tools
Check the model. Scan the source. Contain the agent.
Open source / Alpha
Rowan
Source security analysis for AI applications. Rowan combines Opengrep dataflow, cross-file tracing, dependency analysis, and AI-specific rules in one scan. Findings carry evidence you can inspect.
Private beta
Briar
Run untrusted AI agents and get proof they stayed in. Each run gets its own microVM, one watched way out and no real credentials, and ends with a signed verdict. Built for cyber evals and red-team harnesses.
How Briar works →Open research
Tools you can test against.
Langfail
A deliberately vulnerable MLOps application with 82 planted bugs and a fixed answer key for scanner evaluation.
Explore the benchmark →Quickset
A reproducible corpus for comparing model-file scanners on detection and false positives.
See the project →Framework security
Open LangChain hardening rules and a secure-defaults shim turn threat-model findings into checks developers can run.
See the projects →Get in touch
Talk to the people building the tools.
hello@hedgerow.devQuestions about Rowan, model-file scanning, research, or security disclosures.