Without it
Text on the ledger, like a token's description, can hide instructions aimed at the AI. And an agent's signing key on a server could be copied.
IntegrationsAWS trust layer
Extra safety checks around Pilot, powered by AWS.
AWS adds three separate protections around Harbor. It screens text from the ledger for hidden instructions before the AI reads it, it can offer an AWS-hosted AI model, and it can keep Harbor's signing key locked inside AWS. None of them can approve a trade.
How it works
Watch a request go through it.
The square is the request. Follow it from part to part, read what happens below, and pick an outcome to see what happens when things pass, need you, or stop.
Pick an outcome
AI modelAWS Bedrock
GuardrailsScreens text
KMSHolds the signing key
PilotRequest and receipt
HarborBuilds, applies rules
Your walletSigns exceptions
XRPL TestnetFinal say
Step 01 of 5
Starts when in view
You ask about ledger data
For example, “what is this token?”. Pilot asks Harbor to read it.
Why it matters
Strangers can't slip instructions to the AI through text stored on the ledger.
Why it matters
What changes with it, and what never does.
With it
Guardrails scans that text first and hides anything suspicious. KMS signs inside AWS so the key can't leave. Every receipt shows which screen, model and signer were used.
What never changes
AWS can't grant permissions, raise limits or approve a trade. If a check fails, it stops; it never quietly skips the protection.