Working Terms
A standard mutual NDA should be enough to start.
For early scoping, we prefer practical mutual confidentiality terms that let both sides discuss the system without slowing the work down. Final agreement language can be reviewed by counsel when needed.
Mutual coverage
The NDA should protect confidential information shared by either party, including architecture, source code, prompts, evaluation artifacts, security findings, business context, pricing, and engagement materials.
Confidential information can be written, verbal, visual, technical, operational, or derived from access to systems and materials.
AI-specific material
We expect confidentiality coverage for model behavior, prompt chains, agent traces, RAG sources, MCP/tool configuration, policy rules, eval datasets, reviewer notes, and guardrail findings.
We also expect coverage for vulnerability details, jailbreak paths, privilege issues, and other findings that could be used to abuse a system.
Reasonable exclusions
Standard exclusions should apply for information that is already public, independently developed, rightfully received from another source, or required to be disclosed by law.
The final signed agreement controls. This page is a working posture, not legal advice or a substitute for counsel-reviewed terms.