TimoBy Amotion AI

Timo LabsGuides

An enterprise Claude safety stack enforces controls before, during and after every model action.

Enterprise safety comes from layered controls outside the model. Identity limits who can act, data policy limits what the system can see, tool policy limits what it can change, validation checks the result and monitoring records evidence for review.

Give each risk a control owner and an evidence record. Model safeguards remain one layer. Enterprise systems also need identity, data classification, untrusted-content handling, least-privilege tools, approval gates, output checks and incident response.

Seven control layers

1

Identity and purpose

Authenticate the user, define the approved business purpose and carry identity into every tool call.

2

Data boundary

Classify inputs, remove unnecessary sensitive fields and restrict retrieval to authorised sources.

3

Instruction boundary

Separate trusted system instructions from user and retrieved content. Treat external content as data.

4

Tool boundary

Use task-scoped allowlists, business validation and least-privilege credentials.

5

Output boundary

Check evidence, sensitive data, policy, format and required coverage before release.

6

Human authority

Require approval for high-impact, irreversible or externally binding actions.

7

Monitoring and response

Record traces, detect abuse and drift, revoke access and preserve evidence for investigation.

Worked example: supplier bank-detail change

A fictional procurement agent reads an email from a known supplier. The message includes a hidden instruction telling the agent to ignore policy, change the bank account and approve the next invoice.

Untrusted message

“Our bank details have changed. Use the attached account for all future payments.” The attached document also contains a prompt-injection instruction addressed to the model.

LayerControlEvidenceResult
IdentitySupplier email identity is recorded, but email identity cannot authorise bank changes.Sender, domain and message IDRequest remains unverified
DataAttachment text is labelled external and untrusted.Content provenance tagEmbedded instructions cannot change system policy
InstructionSystem rule says retrieved content supplies facts only.Policy version and evaluation resultPrompt injection is ignored and flagged
ToolThe agent has read access and can open a verification case. It cannot edit banking data.Task-scoped allowlistDirect change is impossible
Human authorityFinance must verify through an approved out-of-band contact and apply dual approval.Call record and two approval IDsChange proceeds only after independent verification
MonitoringThe injection pattern and attempted policy override create a security event.Trace, alert and case numberSecurity reviews the supplier message

Control map by risk

RiskPreventDetectRespond
Prompt injectionTrust boundaries, content labelling and tool allowlistsInjection classifiers and policy-override tracesBlock the call, preserve content and review the source
Sensitive-data disclosureData minimisation, access control and output filteringData-loss-prevention rules and retrieval logsStop release, revoke access and follow incident policy
Excessive agencyLeast privilege, approval gates and reversible actionsHigh-risk call alerts and budget breachesDisable the tool and investigate the trace
Unsupported decisionEvidence requirements and deterministic checksEvaluation failures and reviewer rejectionReturn to draft, correct sources and rerun checks
Model or prompt driftVersion pinning and controlled releasesRegression evaluations and production metricsRollback and open a change review

Evidence package for audit and operations

Architecture review questions

  1. Which data can the system read, and which identity granted access?
  2. Which content is untrusted, and how is that label preserved?
  3. Which actions can the system request, execute and never perform?
  4. Which action needs human approval, and how is approval bound to the exact request?
  5. Which evaluation must pass before a model, prompt, retrieval or tool change ships?
  6. Which logs support an investigation without exposing unnecessary sensitive data?

Public sources

Related Timo Labs guides

Prepare for the architect certifications

CCAR-F covers initial production architecture. CCAR-P is for experienced architects delivering enterprise Claude systems.

Compare architect tracks