Execution authority reviewDecision required
Agent application / privileged action

A tool call is a security decision.
Make it inspectable.

Verify what is running, scan what is entering, constrain what the app may touch, and retain evidence of the decision—before an agent tool becomes an invisible authority channel.

Request tool-risk map →Review the security model
Execution decisionAllowed with limits

wallet.pay

Authority is derived from a verified package and an explicit local grant—not from the model request alone.

Package signature
verified
Binary hash
pinned
Filesystem
/workspace only
Outbound network
deny by default
Audit event #P-1842Policy v4.2
Tool-risk review

Map the authority behind
one dangerous tool call.

Give us a representative action. We will break down its content, package, grant, execution, and audit boundaries against Pilot’s current security surfaces.

What you get
  1. 01Threat-boundary diagram
  2. 02Required grants
  3. 03Recommended control points
Start with one real workflowAbout 2 minutes

Do not include credentials, secrets, or customer data. By submitting, you agree that Pilot may use this information to respond. Privacy Policy.

Control surface

Four checks between intent
and execution.

An agent tool can hold credentials, read files, reach remote services, or launch a process. A convenient configuration is not the same thing as a security boundary. Pilot’s app runtime makes the package, binary, grants, and execution state inspectable, while AEGIS adds a local screening layer for prompt injection and related content risks.

Control 01
Verify

Signed and hash-pinned packages

The catalogue signature is checked and the installed binary hash is revalidated whenever the app launches.

VERIFY
Control 02
Constrain

Explicit capability grants

Filesystem, network, process, signing, and audit access can be declared and reviewed per app.

SCOPE
Control 03
Inspect

Local content screening

AEGIS can scan commands, files, tool results, memory, and skills for injection, impersonation, and obfuscation patterns.

INSPECT
Audit sequence

The trail survives the call.

  1. 01

    Inspect the manifest

    Review the publisher, methods, required grants, protection mode, and integrity state.

  2. 02

    Grant only the required authority

    The local policy stays narrower than the app’s possible capability surface.

  3. 03

    Verify every launch

    Pilot checks the pinned artifact before supervising the app process.

  4. 04

    Retain an execution trail

    App calls and policy-relevant actions can be written to the local audit surface.

Technical basis

Specific product surfaces.
Inspectable in the docs.

Product reviewPilot release 1.13.9Inspect documentation →
  • 01Catalogue public-key verification
  • 02SHA-256 binary pinning
  • 03Per-app grants and guarded mode
  • 04Local AEGIS screening and audit
Adjacent capabilities
Discovery and toolsStop wiring tools one agent at a time. Give capabilities an address.Execution and dataGenerated code should run somewhere disposable. Not on the laptop that hired it.Agent connectivityAgents should know how to reach each other. Not where they happen to run.