Little Canary

Use when you need to screen inbound untrusted text before it reaches a primary model — detecting prompt injection by its behavioral effect on a sacrificial canary model, not just pattern matching — and want a block/flag/pass routing decision plus explicit behavioral-coverage state. Inbound preflight sensor, not a guarantee.

41 stars 9 forksAI, LLMs & Agents

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add hermes-labs-ai/little-canary --skill little-canary
Manual install
git clone --depth 1 https://github.com/hermes-labs-ai/little-canary.git /tmp/little-canary
mkdir -p ~/.claude/skills
cp -r /tmp/little-canary/.agents/skills/little-canary ~/.claude/skills/little-canary

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# little-canary

little-canary detects prompt injection by its effect on a sacrificial canary
model, not just pattern matching: untrusted input hits a powerless model
first, a behavioral check reads the residue, and it returns block, flag, or
pass before your primary model acts. Inbound preflight sensor, not a
guarantee.

## Use it for

- Screening inbound untrusted text before it reaches a main model
- Combining structural pattern checks with sacrificial-canary behavior checks
- Getting a routing decision (block/flag/pass) plus an explicit
  behavioral-coverage state (`REPLAY`, `LIVE`, `MOCK`, `STATIC_ONLY`)
- Running a local HTTP detection server in front of an agent's input path

## Do not use it for

- A guarantee that prompt injection is impossible
- A replacement for runtime containment controls
- A benchmark suite

## Quickstart

```bash
pip install little-canary
little-canary demo --replay
```

Or without installing, via [uv](https://docs.astral.sh/uv/):

```bash
uvx little-canary demo --replay
```

Real output (no packaged replay fixture admitted in this environment):

```
RUN_KIND   REPLAY
MODEL_CALL no — recorded output
CANARY     NOT EXERCISED THIS RUN
EGRESS     none
REPLAY     UNAVAILABLE
DETAIL     no admitted replay fixture is packaged
```

Live contrast against a local Ollama backend:

```bash
uvx little-canary demo --live --backend ollama --model qwen2.5:1.5b
```

## Output shape

- `demo --replay`: verifies analyzer behavior only when admitted response
Read the full skill on GitHub

Details

Skill name
little-canary
Path
.agents/skills/little-canary/SKILL.md
License
Apache-2.0
Repo updated
26 October 2026

FAQ

What does the Little Canary skill do?

Use when you need to screen inbound untrusted text before it reaches a primary model — detecting prompt injection by its behavioral effect on a sacrificial canary model, not just pattern matching — and want a block/flag/pass routing decision plus explicit behavioral-coverage state. Inbound preflight sensor, not a guarantee.

How do I install the Little Canary skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add hermes-labs-ai/little-canary --skill little-canary. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Little Canary popular?

It lives in hermes-labs-ai/little-canary, which has 41 GitHub stars and 9 forks, and is included in 1 curated skill list.

Related skills

Cinetty

51

JosssphZhou/cinetty

Make an animated architecture diagram that plays in the terminal, from one of six finished templates (agent harness loops, a vision model on photos, an agent workflow with a router and workers, a phone-operating agent, a circuit or service board with signals on wires). Use when the user asks for an architecture animation, an animated flow chart, a terminal video or a demo clip of how their system works, in English or Chinese (for example 架构动画, 终端流程图, 给项目做一支动画). The skill reads the user's repository, fills one project.yaml, checks it until it passes and plays it in the current terminal. Do not

AI, LLMs & Agents10

Cleopatra Modern Life

52

10291124jcjc-dotcom/cleopatra-modern-life

古人的现代日常:把一个古代人物(埃及女王、罗马角斗士、维京首领……)放进现代日常,做成一条 30 多秒、全程一本正经的 AI 短视频。从拆参考视频、写梗(古代场景 + 现代物件 + 古法对照物 + 结果)、出一张人物参考图、写 3 机位的即梦 Seedance 提示词,到检查生成的片段、测歌找高潮、按节拍自动剪辑出 4:3 / 9:16 两个版本。用户说「做一期 XX 的现代日常」「古人过现代生活」「历史人物做现代事」「埃及女王那种视频」,或者要改已有一期的梗、提示词、剪辑时使用。Make an "ancient person, modern life" AI short end to end: gags, one character reference, multi-angle Seedance prompts, clip review, and a beat-synced edit.

AI, LLMs & Agents11

Animated Short Film

43

sonlovinbot-team/Muse-animated-short-film

Skill dựng phim hoạt hình ngắn từ đầu đến cuối trên Muse AI: thu thập yêu cầu (thời lượng, ngôn ngữ, tỷ lệ khung hình), vẽ character sheets, chốt kịch bản story bible, keyframes, mega prompt từng cảnh, generate video, làm VO, dựng phim bằng ffmpeg. Tác giả: Đặng Hữu Sơn. Dùng khi user muốn làm phim hoạt hình ngắn từ con số 0 trên Muse AI.

AI, LLMs & Agents24

Manage Skills

43

umutbozdag/agent-skills-manager

Discover, list, create, edit, toggle, copy, move, and delete AI agent skills across 11 tools (Cursor, Claude, Agents, Windsurf, Copilot, Codex, Cline, Aider, Continue, Roo Code, Augment)

AI, LLMs & AgentsIn 1 list

Prompt Refine

52

Apeironics/prompt-refine-skill

Silently restructures the user's natural-language prompt into the format the model CURRENTLY running this skill handles best, then answers. On activation it identifies which model family is executing it (Claude, GPT, Gemini, Llama, DeepSeek, Mistral, Qwen, Grok, Perplexity, Kimi, GLM, Command, Nova, or Phi) and loads that one model's official strategy — so the optimization always matches the model that actually runs it. Activate with /prompt-refine. Use when users want better answers without learning prompt engineering.

AI, LLMs & Agents

Im Human

56

chang416/im-human

台灣繁體中文與英文的語態控制與去 AI 味編輯技能,兩種模式。語態模式:使用者說「我是人」「開啟我是人」「說人話模式」或要 Claude「用我是人講話」時啟動,之後 Claude 自己的每一句輸出都改用台灣人日常的繁體中文(英文則像母語者隨手寫的),禁用 AI 常見的開場白、路標詞、空泛拔高、翻譯腔與簡報式排版,直到使用者說關閉。編輯模式:使用者說「去 AI 味」「說人話」「改自然」「不要像模板」「幫我審稿」,或要改寫貼文、文案、Email、文章、技術文件、小說時啟動,在不更動事實、數字、引用與作者立場的前提下清掉模板感。Also handles English requests to strip AI writing style or make text read like a human wrote it. 只做台灣繁體中文與英文,不支援簡體中文。不用於翻譯、事實查核、憑空創作,也不用於規避 AI 偵測器。

AI, LLMs & Agents10

Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.