Security and trust boundaries
Apply AletheIA security guidance through explicit trust boundaries, advisory checks, and meaningful human review.
Security in AletheIA begins by asking a practical question: what can this Work Slice trust, access, change, disclose, or influence?
The answer determines the context you load, the checks you apply, the proof you require, and when a human must decide. It does not turn guidance into automatic enforcement.
Advisory guidance Human-gated Source-backed evidenceDefinition
A trust boundary is the point where content, authority, identity, permission, or responsibility changes. Examples include:
- a user-provided file entering an AI-assisted workflow;
- data moving from one tenant, client, or organization to another;
- an agent receiving a tool permission;
- an external instruction attempting to influence a governed task;
- a public artifact leaving a private project context.
Why it matters
AI-assisted work can make unfamiliar input appear useful or authoritative. AletheIA treats external content as data to be evaluated—not as instructions to follow—and preserves uncertainty where evidence is incomplete.
How it works
Identify the boundary
Record which inputs, identities, tenants, services, tools, or publication surfaces cross a trust boundary. Do not assume a source is safe because it is readable or well formatted.
Classify the context
Use the canonical sensitivity vocabulary—public, internal, confidential, restricted, or regulated—and keep uncertain classification visible.
Apply the relevant guidance
Use the AI Agent Security and Web App Security packs when the Work Slice actually involves agents, prompt injection, untrusted content, web/API behavior, authentication, tenant boundaries, or secrets.
Pause for meaningful human review
Escalate when a mandatory review condition applies: high-impact decisions, sensitivity uncertainty, external publication, client delivery, source conflict, regulatory risk, or a trust-boundary crossing.
Preserve evidence and close honestly
Record the source references, what was checked, the human decision when needed, and any unavailable proof. Do not claim technical enforcement that the real environment did not provide.
Practical example
A Work Slice receives a customer-provided support export and asks an AI runtime to summarize it for a public product update.
| Question | Safe posture |
|---|---|
| Is the export an instruction source? | No. Treat it as untrusted data. |
| Does it cross a boundary? | Yes. Customer information may enter a public-facing decision. |
| What guidance applies? | Restricted-context, source evaluation, data leakage, and human-review guidance. |
| Can the runtime publish? | No. AletheIA does not grant that authority. |
| What counts as proof? | Classified source references, review decision, and validated final artifact—not an assumed safe prompt. |
Delivered guidance and current limits
AI Agent Security and Prompt Injection
Delivered as a manual/advisory pack for real Work Slices involving agent inputs, prompts, tools, or untrusted content.
Web App Security and Trust Boundaries
Delivered as a manual/advisory pack for real web, API, authentication, tenant, secret, or trust-boundary work.
Related concepts
- Human-review criteria
- Prompt injection in sources checklist
- Restricted knowledge usage policy
- Domain governance packs
- Runtime and safety in the ecosystem