Gong et al.: Authority–Inference Separation in Agentic Finance (Systems)

Evidence record

October 20 revised journal article; full-text fetching hit HTTP 429. Official indexed abstract and operator-supplied excerpt were available, but full journal text and replication were not independently inspected.

Read the original source ↗

Claim-level citations (3)

• supportsAgent-wallet tests blocked transaction effects despite successful prompt injection: A gateway that authorized a proposal but did not bind the authorized decision to execution bytes admitted 20/20 targeted post-decision modifications in the authors’ separate synthetic fixtures.

  Journal authorization gateway comparison, operator supplied metrics; full text not independently retrieved

• supportsAgent-wallet tests blocked transaction effects despite successful prompt injection: Only 6 of 16 legitimate model-produced requests were authorized in the reported model-in-the-loop experiments, limiting availability in that setup.

  Journal model-in-the-loop legitimate-request acceptance, operator supplied metrics

• supportsAgent-wallet tests blocked transaction effects despite successful prompt injection: The revised October study reports eight of eight prompt-injection attempts successfully changed proposed transaction counterparties in two evaluated open-weight models, with zero unauthorized executions under the independent authorization layer.

  Journal October 20 model-in-loop result, as presented in indexed article and operator-supplied research material