control-concepts
Guide to Control Tower's AI control concepts. Use when the user asks about control protocols, blue team and red team roles, usefulness, safety, audit budgets, trusted monitoring, or the conceptual model behind Control Tower.
- License 许可
- MIT
- 最近更新
- 2026-11-13
- Stars
- ★ 46
- Forks
- 53
01 前言 Front Matter
- name
- control-concepts
- description
- Guide to Control Tower's AI control concepts. Use when the user asks about control protocols, blue team and red team roles, usefulness, safety, audit budgets, trusted monitoring, or the conceptual model behind Control Tower.
02 资源文件 1
- 参考文档README.md164 B
03 SKILL.md 正文
@docs/control-concepts.md
04 同仓库更多 Skills
validate-straj
Validate a sabotage trajectory (straj) end-to-end before trusting it as a keeper. Reconstructs the attack mechanism, checks for blatant reward-hacking and for unearned stealth (an environment flaw that made the side task trivially easy and low-suspicion), replays the side-task scorer in a fresh sandbox, and re-scores the replay with the `simple` (action) + `trajectory` (traj) monitors on opus-4-8 and gemini-3.1-pro (scored once per model, escalating a flagged monitor to a majority of 3). Use when the user wants to validate / verify / QA strajs, confirm a straj still evades the canonical monitor, or check a "clean keeper" list of trajectories.
trajectories
Guide for working with Control Tower trajectories. Use when the user asks about trajectory formats, trajectory IDs, converting .eval logs, uploading or downloading trajectories, replaying trajectories, or trajectory CLI commands.
safety-curves
Guide for understanding and explaining Control Tower safety curves. Use when the user asks about safety-usefulness tradeoffs, audit budgets, attack success rates, monitor thresholds, safety metrics, or interpreting safety curve plots.
running-evals
Guide for running evaluations in Control Tower. Use when the user asks about running evals, ct run eval, policies (honest, attack, test, replay), protocols, task sets, uploading results, replaying trajectories, or evaluation options. IMPORTANT - Use this skill instead of the MCP tools for guidance on running evaluations.
monitoring
Guide for monitoring trajectories in Control Tower. Use when the user asks about monitoring, suspicion scores, the ct run monitor command, monitor prompts and fields, or trajectory analysis. IMPORTANT - Use this skill instead of the MCP tools for guidance on monitoring trajectories.
making-strajs
Guide for creating sabotage trajectories (strajs) in Control Tower. Use when the user asks about creating strajs, monitor evasion, attack strategies, straj documentation, or trajectory development workflows.