Security & trust

Propose capped shadow prices without changing live prices

Estimate SLA failure risk from bounded outcome and latency samples

Identify capability shortages from verified demand and provider supply

Deterministic counterparty policy decision using supplied reputation evidence and transaction limits

Inspect untrusted tool output before it enters an agent context or triggers another action

Validate least-privilege capabilities, data scope, destinations, budget and expiry before agent-to-agent delegation

Screen persistent agent-memory writes for poisoning, missing provenance, secret requests and unsafe instructions

Inspect an MCP tool manifest before installation or trust, including collisions, missing schemas and instruction injection

Enforce declared per-call, task, session and period budgets before execution

Detect safe replays and conflicting reuse of an idempotency key

Determine whether an agent action may run, needs approval or must be denied

Validate HTTPS callback and redirect destinations against an explicit host allowlist

Detect changes between a trusted tool baseline and its current manifest

Bind declared MCP domain, endpoint, manifest, key and version in signed evidence

Hash and label agent context by declared integrity, confidentiality and source

Compare multiple agent results and sign exact agreement or conflict evidence

Calculate and sign availability and latency evidence from bounded samples

Validate declared signature status, freshness and payload evidence

Check declared dependency identity, source and digest coverage

Select among conflicting results using explicit policy

Calculate data age against an explicit freshness policy

Attest declared DNS and HTTPS domain-control proof status

Bind declared consent to action purpose, scope and expiry

Plan bounded idempotent resumption from a checkpoint

Screen a declared subject against independently retrieved official sanctions sources and return traceable evidence

Assemble verified failure evidence and recommend review without executing refunds

Commit request, authority, execution, payment and delivery evidence in one auditable proof

Map mandate fields for review without claiming AP2 authorization or supported production control

Fail closed on inactive mandates and per-transaction or remaining-budget violations

Aggregate diverse verified observations with mandatory low-sample abstention