Unverified50% confidenceFactExact time
ScopeBench is a benchmark containing 30 'dead-end' agentic security tasks whose objectives can only be achieved by violating the engagement scope.
1
Sources
50%
Confidence
Long-term
Relevance
9/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEach task in ScopeBench is designed with two control conditions: the Scopeless condition measures raw capability, and the Scoped condition measures compliance capability, with both sharing the same environment, validator, and objective.72% similarUnverifiedGoal mode embeds circuit-breaker logic within the task execution framework, proactively interrupting the process when anomalous signals such as classification failure rate exceed a threshold, thereby preventing cascading failures.64% similarUnverified/goal设计了三层防护机制：零工具调用检测、预算控制、完成审计协议，防止Agent空转58% similarUnverifiedThe core idea of the TIP attack is to disguise harmful objectives as seemingly benign technical tasks, inducing the model to produce content that should have been blocked by safety mechanisms.58% similarUnverifiedTask horizon refers to the number of operational steps a model can independently complete without seeking human feedback57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/955460API
curl https://kongchang.com/api/v1/knowledge/claims/955460MCP
get_claim(id=955460)