Shenzhen University and HKUST propose ECA: evidence checks before agent actions
jiqizhixin · x · 2026-09-22
Shenzhen University and HKUST present ECA (Evidence-Carrying Multimodal Agents), addressing how one false premise ("the address is confirmed") can derail an entire agent task chain. ECA separates duties: a multimodal model understands pages/screenshots and proposes tool calls, while independent verifiers extract evidence from the original content — confirming buttons exist and match the task, and validating recipients, intent, and attachment permissions before sending. Tools execute only after evidence passes inspection, treating hallucination as an exploit rather than a text-generation problem.
More from Safety
- Agent product Instinct admits leaking another user's financial data into a chat — Scobleizer · 2026-09-22
- AI hallucination of nuclear cargo on Chinese ship nearly triggered US military strike — CurieuxExplorer · 2026-09-22
- Feeld dating app bugs let anyone read chats and access private photos — NathanpmYoung · 2026-09-22
- WSJ op-ed: an AI antitrust exemption would invite collusion; safety collaboration is already legal — AlexTensor · 2026-09-22
- Hinton: US and China won't agree on election deepfakes — every government wants the option — victor_explore · 2026-09-22
- Dev asks why NIST CSF 2.0 and AI RMF 1.0 went unused before new AI rules — AlexTensor · 2026-09-22