Can we allow this agent, with these flags?
A version-pinned allow/deny readout for AppSec and AI-platform teams. One named agent. An exact release or commit. The configuration you plan to roll out.
Who this is for
You are choosing an agent, expanding its permissions, changing its configuration, or approving a version update for your engineers. Your AppSec, product-security, or AI-platform owner needs to know whether the intended controls match the recorded behavior.
The starting point is a named agent and a concrete deployment decision. We agree on the available evidence and what it can establish before accepting the scope.
The deliverable
Observed behavior
- Executes without asking, boundary held, or not established.
- The version and configuration behind each observation.
- Matched controls and normal-task completion.
- Evidence level, source references, and coverage gaps.
Intended policy
- Your allow, ask, and deny decisions per capability.
- Mismatches between intent and observation.
- Deliberately accepted behavior and unused permissions.
- A private readout and documented follow-up.
Current scope
This offer covers a review of recorded, controlled evidence and configuration. Live reproduction against a real agent image is not included. New execution work requires its own agreed scope and access plan.
A held boundary is a useful result. So is a clearly identified gap in coverage. We do not promise findings, a severity rating, a certification, or permission to deploy beyond the evidence reviewed.
Customer evidence and findings stay private. Any case study, reference, or disclosure requires a separate agreement.
Evidence you can inspect
Reports distinguish attempted actions, recorded executions, and confirmed effects. A normal-task control must complete successfully to support a clean comparison. Errors and incomplete runs remain visible.
Every conclusion binds to the named version, snapshot, and flags. A version or configuration change needs a new review of the affected evidence.
Name the agent, version, and flags.
Bring the rollout decision, intended permissions, evidence available, and timing. We confirm the scope, fee, delivery date, NDA needs, and stop conditions in writing before work begins.
Use synthetic data and an agreed test environment. Do not send production credentials, customer data, or access to engineer endpoints.
Opens your email application. No information is submitted by this page.