用 Jev 的 Noul/Score/Choice 检测 agent 漂移与提示注入
安全研究者在演讲中展示用 Jev 分类 agent 是否跑偏、被注入攻击，并做了分析 Codex/Claude Code 会话的实用工具。
🤓 At @UnpromptedAU I presented some of the experimentation I did with @typesafeai Jev. I used Jev to evaluate AI agent drift and misalignment to detect if my agent is going rogue or being compromised. Jev is a model used for data classification and decision-making. It uses 3 different primitives: Noul: Is the agent off-task? Score: How severe is the drift? Choice: What behavior is it? You can define your own criteria depending on the data you are analyzing. It is cheaper and faster than an LLM, and the data you classify can be anything, which makes it great for threat intel and cybersecurity! In my example below, Jev is evaluating if my agent is being compromised through a prompt injection loaded from external data and passed into the context of my agent. I also created a utility to analyze my codex and claude code agent sessions and discover if something went wrong or was unusual during the session. I can see a lot of opportunities for threat detection and data classification. I still need to experiment more but it opens new possibilities. More to come 🧐