Unverified50% confidenceFactExact time
τ-bench's criterion is: even if the agent successfully helps the user refund a ticket, as long as the refund action violates policy, the task is still judged as failed, i.e., it measures "task completion under the premise of following the rules".
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIf the agent refuses the malicious call but also fails to complete the task it should have completed, this is essentially a usability loss rather than a security success.72% similarUnverifiedError cascading refers to a downstream agent treating the erroneous output of one agent as trusted input for further processing; one prevention method is to allow downstream agents to proactively refuse uncertain conclusions (e.g., triggering a human fallback when returned data is insufficient).72% similarUnverifiedIf the agent fails to complete the task in both the benign and malicious versions, it indicates that its refusal was likely just confused by the injected text and not a genuine security judgment.71% similarUnverified作者认为执行轨迹（trace）比自动评判结论更可靠，是判断Agent失败是否真实存在的关键依据71% similarUnverifiedThere is a risk that a single Agent handling tasks may go off track (hallucinate), potentially misjudging an order and issuing an erroneous full refund.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1007500API
curl https://kongchang.com/api/v1/knowledge/claims/1007500MCP
get_claim(id=1007500)