Microsoft 开源了一个东西,把 AI 智能体的安全测试做成 pytest:RAMPART(建在 PyRIT 上)。 对抗场景写成测试 → 在 CI 里 pass/fail。 可以定「1000 次重跑里至少 80% 安全」这种阈值。 #AI #AIagent

Microsoft 5 月 20 号开源了一个东西,把 AI 智能体的安全测试做成了 pytest:RAMPART。 写一个 prompt-injection 场景 → 跑成测试用例 → 在 CI 里给 pass/fail。统计模式可以定「至少 80% 跑次安全」。 Agent 安全这块,正式进入 SRE 工具链了。 #AI #AIagent

GitHub's internal repos were breached after an employee installed a malicious VSCode extension. 3,800 repos exposed. Your dev toolchain is an attack surface. If you're shipping AI agents, test your…

dug into microsoft's new RAMPART repo. why does an agent security tool need 40+ files and 5 __init__.py files just to re-export the same functions? 20% features, 80% ceremony. the core ideas are…

Microsoft open-sourced a pytest framework for testing AI agent safety. RAMPART: adversarial attacks, benign failures, harm assertions — all as pytest cases. Runs in CI with your other tests. →

These tools cover two parts of the safety workflow: - Clarity helps teams question the design before real implementation - RAMPART helps teams continuously test whether the system behaves safely as…