ouros-ai-security

Audit Ouros AI, MIDAS, MCP and tool-calling security. Use when reviewing prompt-injection resistance, tool authorization, cross-user/thread isolation, MCP boundaries, guardrails, model-controlled argu

SKILL.md

GitHub

Ouros AI Security

Treat model output as untrusted input to privileged systems.

Read ../../references/ouros-security-model.md for the Ouros AI/tool trust path.

Security invariants

Verify these before chasing prompt cleverness:

  1. A model cannot grant itself permissions.
  2. Every privileged tool re-checks identity/authorization server-side.
  3. User-controlled or model-generated resource IDs are checked against the authenticated principal.
  4. Thread/history retrieval is principal-bound.
  5. Tool responses expose only data the caller is authorized to receive.
  6. Hidden/system instructions and secrets are not relied upon as the sole authorization boundary.
  7. Guardrail failure does not become authorization failure.

Audit order

A. Tool authorization

Map each tool to required identity, role/scope, accepted resource identifiers, mutating vs read-only behavior, and server-side ownership checks.

Flag any tool where authorization depends primarily on the model choosing a safe argument.

B. Cross-principal isolation

Using dedicated test principals, check whether thread IDs, farm/enterprise/resource IDs, cached context, or memory can cross ownership boundaries.

Use ../safe-runtime-testing/SKILL.md for runtime probes.

C. Prompt and indirect-injection boundary

Test whether untrusted content can influence tool selection or arguments beyond the caller's existing permissions.

The key question is not whether the model followed an instruction, but whether following it crossed a protected boundary.

D. Data minimization

Review model context assembly, tool outputs, logs/traces, error messages, and telemetry for unnecessary secrets, identifiers, or personal data.

E. Guardrails

Evaluate guardrails as policy/usability controls, not substitutes for authentication or authorization.

A guardrail bypass is security-relevant only when it produces concrete protected impact or violates an explicit safety requirement.

Finding quality

Do not report vague claims such as prompt injection possible. Report the violated invariant, protected boundary, prerequisites, and observable impact.

Pass every candidate to ../finding-verifier/SKILL.md.