Skip to content

v0.9.0 - Executable prompt-injection reference eval

Choose a tag to compare

View all tags
@lindixu6-hash lindixu6-hash released this 16 Aug 21:51
v0.9.0
5167473

What shipped

  • Deterministic reference Agent runtime with separate trusted-instruction and source-labeled untrusted-content channels.
  • Executes all eight malicious and benign prompt-injection fixtures instead of hand-authoring result files.
  • External evaluator emits v0.7 Eval Results plus response, assertion, tool-trace, and policy-trace artifacts.
  • Failure-path regression proves an observed/expected mismatch produces a failing result.
  • CI validates generated results with public @v0 and retains 34 evidence files as the reference-eval-evidence artifact.
  • English and Simplified Chinese integration guidance explicitly limits the result to this deterministic reference policy; it is not a model benchmark or Content OS safety claim.
  • 94 deterministic tests.

中文说明

  • 新增确定性参考 Agent Runtime，可信指令与带来源标签的不可信内容走独立通道。
  • 真实执行全部 8 条恶意与良性 Fixture，不手写结果文件冒充执行。
  • 外部 Evaluator 生成 v0.7 Eval Result、回答、断言、工具 Trace 与策略 Trace。
  • 失败路径回归测试证明实际结果与预期不一致时会产出 fail。
  • CI 使用公开 @v0 校验生成结果，并将 34 个证据文件保留为 Artifact。
  • 中英文文档明确：这只证明确定性参考策略行为，不是模型 Benchmark，也不是 Content OS 安全证明。
  • 共 94 项确定性测试。

CI evidence: https://github.com/lindixu6-hash/awesome-agentic-engineering/actions/runs/31974669422

Closes #13.

Assets 2
Loading