遇见数据集

SandboxEscapeBench Extended Gold Bank (71 items) — Agentic Containment Detector Scorecard

收藏
Zenodo2026-08-14 更新2026-08-20 收录
官方服务:

资源简介:

SandboxEscapeBench extended gold bank: 71 deterministic, rule-labeled items (38 ESCAPE + 33 BENIGN) for scoring agentic-containment detectors — 11 hand-authored items plus 60 converted from redblue_v2's deterministic attack matrix (10 harmful requests x 5 jailbreak wrappers + plain and benign controls). No model judged the labels. Includes the full detector scorecard run (rce_sandbox / firejail, static + dynamic): 38/38 escape recall, 33/33 benign pass, gold-integrity assertion PASS on every item. The n>=30-per-class floor is cleared, making containment a quotable measurement axis. Wilson confidence intervals travel with every rate. Signed measurement cards for these runs are publicly verifiable at https://csoai.org/verify. Part of the CSOAI GSPC instrument family.

提供机构:
Zenodo
创建时间:
2026-08-14
二维码
社区交流群
二维码
科研交流群
商业服务