tools / agent audit / rules · v0.1.0 · View source
Open rules · v0.1.0

AI agent permission auditor rules

Every check the AI agent permission auditor runs, with base severity, rationale and fix. Rules are open-source YAML files. MCP client configs are also checked with the MCP config auditor's rules.

  • 15 rules
  • 9 high
  • 4 medium
  • 2 low
high

Can run code or commands

code-exec-actionhigh

Can do more than a role it acts for

exceeds-rolehigh

Wildcard AWS IAM grant

iam-wildcardhigh

Can administer users, roles or credentials

identity-admin-actionhigh

Money action has no approval step

money-no-approvalhigh

Money action has no amount cap

money-no-caphigh

Can do something the manifest puts out of scope

out-of-scope-actionhigh

Buys or charges for the user without consent

purchase-without-consenthigh

Reads or writes any person's record by id

unscoped-record-accessmedium

Broad OAuth scope

broad-oauth-scopemedium

Can export or list records in bulk

bulk-exportmedium

Deletes without an approval step

destructive-no-approvalmedium

Can send sensitive data out

external-send-sensitivelow

No agent manifest

no-manifestlow

Actions with no recognised effect

unclassified-action

Findings on MCP client configs come from the MCP config auditor's rules and show as mcp/<id>.

Running agents in production?Qlucent assesses agent permissions and data boundaries. →