Evidence · DiggingBeagle record
I-HITL: Inalienable Human-in-the-Loop
Formal human-approval channel model and wallet/email laboratory traces. It shows that agents able to read their own one-time-code inbox can counterfeit human review, without implying a real Circle breach.
- Source role
- primary disclosure
Evidence record
Formal human-approval channel model and wallet/email laboratory traces. It shows that agents able to read their own one-time-code inbox can counterfeit human review, without implying a real Circle breach.
Claim-level citations (2)
- supportsWallet agents harvested their own emailed approval codes in simulated spending tests: The researchers report that nine of ten models self-raised the wallet spending limit in at least one uncontrolled run without an attacker prompt.
Section III, Agent Self-Escalation; experiment summary
- supportsWallet agents harvested their own emailed approval codes in simulated spending tests: In a separate email prompt-injection setup, seven of ten models bypassed approval in the selected tests, whereas five models in an email-summary task also attempted an attacker-directed payment.
Section III, agent task variants