Agent boundary evaluation
Examine unauthorized requests and adversarial instructions. Keep a model’s refusal separate from evidence that an execution guard rejected a proposal.
AI Agent Testing and Evaluation
Clear boundaries.
Evidence you
can examine.
We’re developing practical evaluations of what AI agents are allowed to do—and what happens when those boundaries are challenged.
Explore the prototype report ↗See our approachScoped testing. Preserved findings. Clear limitations. Independent review is being sought; accreditation has not been obtained.
Permission boundariesControlled software simulationReviewable evidenceExplicit limitations
Proposed services
Understand the boundary
before relying on the agent.
Our initial focus is a defined agent configuration: its tools, permissions and behavior in controlled tests. Services remain under development.
Execution control testing
Exercise allowed and denied actions, argument validation, permission revocation and state changes in a software simulator.
Findings and review preparation
Document scope, test coverage, records and unresolved findings so a qualified reviewer can inspect the conclusions and reproduce the checks.
Our proposed approach
From an agreed scope
to a traceable conclusion.
Define
Specify the model, tools, permissions, environment and acceptance criteria before testing.
Challenge
Use controlled cases to exercise the declared boundaries and observe failures.
Preserve
Retain proposals, decisions, errors and state changes, including failed and incomplete trials.
Review
Explain what the evidence supports, what it leaves unresolved and what requires retesting.
Prototype evidence
Read the result.
Then read its limits.
A small developer-run demonstration of the guarded simulator. These observations are not an independent assessment or operational certificate.
Control demonstration · October 16 2026
Permissions and evidence handling
6Scripted guard scenarios passed
2Evidence change checks passed
0Live model calls in this demo
What the demonstration checked
| Scenario | Observed outcome |
|---|---|
| Denied unlock / unknown tool | Blocked; state unchanged |
| Malformed JSON / permission spoofing | Arguments rejected |
| Navigation duplicate and cancellation | Duplicate blocked; cancellation applied |
| Permission revoked before next action | Later request blocked |
| Captured evidence edited or removed | Digest changed |
What these results do not establish
No real vehicle, live model, client data or independent reviewer was involved. Rejecting an unknown “disable monitor” tool does not establish production-monitor security. Digest comparisons detect changes only when the original digest is trusted; they are not authenticated or tamper-proof storage. Human alert delivery and privacy isolation were not tested.
Current readiness and remaining work
Five separate live authorization trials produced explicit refusals without tool proposals. They do not exercise the execution guard. Remaining work includes missing injection trials, held-out evaluations, privacy fixtures, stronger provenance and independent review. No accreditation or operational certificate has been issued.
This summary reflects the recorded prototype checkpoint. It does not update automatically from ongoing tests.
About Steward
Accountability starts
with accurate claims.
Steward AI Assurance is a business in development led by Luisito Espanola. We’re building a focused approach to AI-agent permissions, execution controls and evidence reporting.
Business registration is pending. External technical review is being sought. We do not claim accreditation, government approval or a guarantee that any AI system is safe.
A focused starting point
Let’s define the question
before promising an answer.
For a future evaluation, we would first discuss the agent configuration, permitted tools, testing environment, data sensitivity and the evidence you need.
Business inquiries: info@stewardaiassurance.com
Please start with a brief description. Do not email credentials, private datasets or sensitive client information. Paid engagements are not yet available.