Least-privilege tool access
Every tool / API the agent can invoke is documented, justified, and scoped to the minimum permission needed for the documented business task.
Full guidance is available with a trial workspace.
Public preview
AgentProof Intelligence — a public taste of the source-backed library
AgentProof references named, public sources from Microsoft, NIST, ISO/IEC, OWASP, the EU AI Act, Google SAIF, the Cloud Security Alliance, and MITRE ATLAS. AgentProof summarises and maps the guidance — it never copies full standards text and never claims certification or legal compliance.
Access level: public_preview · public preview never exposes the full library, full evidence examples, full framework mappings, full radar, or pack history.
Clear intended use
A short, written intended-use statement names what the agent is for, who it serves, and what it must NOT do.
Full guidance is available with a trial workspace.
Human approval for high-impact actions
Every action with financial, legal, customer-impact, or safety consequence requires named human approval before execution.
Full guidance is available with a trial workspace.
Audit trail
Every agent decision and action is recorded with timestamp, input, output, and acting identity.
Full guidance is available with a trial workspace.
Continuous reassessment
The agent is re-reviewed on a documented cadence + after any material change.
Full guidance is available with a trial workspace.
Risky patterns to avoid
Excessive permissions
Agent has broader tool / data access than its intended use requires.
Broad autonomous action without approval
Agent can take material action without a human approval step.
No audit trail
Agent decisions and actions are not recorded.
No intended-use boundary
Agent has no documented intended use; scope creep is likely.
Prompt / tool injection exposure not considered
Agent has no documented protection or testing against prompt injection.
Sample controls (5 of 20)
Intended-use statement
Make the agent's purpose and limits unambiguous.
Tool permission register
Enforce least-privilege tool access.
Human approval rule
Gate high-impact actions behind named human approval.
Audit log
Record agent decisions and actions for accountability.
Monitoring / review cadence
Keep the agent observed and reviewed on a documented cadence.
Go deeper in the Library
Want to read the framework first?
Public preview is free. The full library is available inside the trial workspace.