The report
Financial settlement verification is supposed to connect a claimed repayment to an independently observed transfer. The Priors security advisory describes how an authorized credit writer could supply an empty list of transfer anchors and still obtain an apparently successful check, with anchorsChecked equal to zero. The problem is not a misbehaving language model by itself. It is a verifier accepting the absence of required evidence as proof. Empty-anchor finding.
The advisory also identifies a separate way to distort a financial record: a writer-influenced event mapping could change the interpretation of repayment and default fields. That creates an apparent accounting result without requiring the transfer history to substantiate that classification. Both paths begin with a writer that can submit the relevant data, and the demonstration was conducted with simulated records. Event-schema finding.
The maintainer reports remediation at commit c3c9b19. The disclosed material does not establish an actual victim, a production-scale loss, or whether every deployed ledger independently reconciled the purported credit with verified token movement. Maintainer advisory.
A robust settlement boundary has three independent checks. There must be at least one required, valid evidence anchor; each event must be decoded using a pinned contract identity and ABI rather than a writer-provided interpretation; and the financial state change must reconcile with actual token amounts and decimals. An AI agent may prepare the settlement request, but it cannot confer independent verification authority on its own output. These checks target the two demonstrated mechanisms without assuming that all financial-agent failures share one cause.