How FLINT Verifies AI Agent Authority

What Is AI Agent Authority Verification?

AI agent authority verification determines whether a specific agent may perform a specific action for a specific principal at a specific moment. It is narrower than general identity verification and more actionable than a risk score because it evaluates the requested transaction against delegated authority before execution.

FLINT screens agents with or without a passport. It combines independent evidence about identity, authority, payment context, environment, and reputation, then returns a verdict and signed verification record. The record lets both sides verify what FLINT checked, what was unavailable, and what decision it issued.

Principal Identity

Principal Identity asks who authorized or controls the agent. It considers authority assertions, credential attestations, delegation context, and the relationship between the principal and the requested transaction.

Agent Identity

Agent Identity asks which non-human actor is making the request and whether its presented identity is consistent. It considers agent identifiers, signed identity material, delegation context, and continuity observations that connect the request to the acting agent.

Wallet Provenance

Wallet Provenance asks whether the payment instrument and counterparty context support the transaction. It considers protocol-native transaction fields, wallet history and provenance categories, sanctions screening results, and network reputation.

Authorization Scope

Authorization Scope asks whether the observed action fits the authority the principal declared. It compares the requested amount, counterparty, action, and timing context with the scope presented for the transaction.

Environment Identity

Environment Identity asks whether the agent is operating in a consistent and credible runtime context. It considers device and environment intelligence, session continuity, workload context, and network reputation without exposing raw provider data to counterparties.

Cross-Merchant Reputation

Cross-Merchant Reputation asks what prior commercial outcomes say about the agent. It considers completed and disputed transactions, mutual acknowledgments, counterparty diversity, inactivity, and reputation connected to a durable passport when one exists.

What ALLOW, STEP-UP, REVIEW, and BLOCK Mean

The caller receives 1 of 4 decisions and can act on it before payment or fulfillment.

• ALLOW The request can proceed under the authority and transaction context FLINT evaluated.
• STEP-UP The request needs another verifiable action before FLINT can make a new decision.
• REVIEW The request should pause for human review because the available evidence does not support automated execution.
• BLOCK The request should stop because affirmative evidence does not support execution.

FLINT issues the decision. The integrating wallet, API, checkout, or fulfillment system controls enforcement. A verification record cannot hold, cancel, refund, or release value on its own.

How FLINT Treats Missing Evidence

FLINT reports a check as performed only when the relevant screening capability genuinely answers. A timeout, absent input, unsupported transaction type, or provider failure is labeled unavailable or not applicable. Missing evidence is never silently converted into a clean result.

Evidence completeness and risk are different dimensions. An incomplete check can require STEP-UP without implying fraud. A confirmed sanctions match, failed authority proof, tampering signal, or mandate violation is affirmative adverse evidence and can support BLOCK under the applicable rules version.

How STEP-UP resolves

STEP-UP is not a dead end. The response includes a machine-readable menu of cures that attach to the agent and carry forward to future verifications.

• present_passport Present an existing passport so the next check can use its identity and reputation context.
• mint_passport Mint and claim a free Agent Passport where reputation can accrue for the life of the passport-agent pairing.
• declare_authorization_scope Declare the amount, counterparty, action, and timing boundaries the agent is authorized to use.
• attach_principal_attestation Attach verifiable evidence of the principal who authorized or controls the agent.
• principal_reauthorization Obtain fresh authority from the principal for the requested action.
• single_use_token Provide authority limited to the current transaction.
• reduce_amount Lower the amount and submit the changed transaction for a fresh decision.
• staged_amounts Divide the transaction into smaller stages that can be evaluated separately.
• escrow_route Use an escrow arrangement before either side releases value.
• mutual_acknowledgment_precommit Have both sides commit to acknowledging the completed transaction before execution.

How Agent Reputation Persists

Reputation can accrue from completed, disputed, refunded, and mutually acknowledged outcomes. It should consider recency, counterparty diversity, and the quality of the evidence connecting each outcome to the same agent.

A durable passport gives reputation a stable attachment point across restarts and redeploys. It does not create permanent authority. Every consequential action still needs a fresh check against the current mandate, transaction, and runtime evidence. Agents without passports are still screened, but they have less durable context to carry forward.

What a Signed Verification Record Proves

Every decision produces a signed verification record. FLINT signs the record with ES256 plus a post-quantum signature, and either party can independently check its integrity against published verification material. The buyer and seller can retain the same record and link their acknowledgments or outcomes to it.

The signature proves that the record has not been altered since FLINT issued it. The record proves what FLINT evaluated and decided. It does not prove that payment settled, that fulfillment occurred, or that the transaction later produced a legitimate outcome unless those events are separately captured and linked.

What FLINT Does Not Publish

FLINT publishes evidence categories, verdict semantics, record structure, and verification methods. It does not publish exact thresholds, signal weights, evasion-sensitive rules, or private provider data. That boundary lets buyers and sellers understand the decision without giving attackers a tuning guide.

Common Questions