Little Canary

Screen untrusted prompts for signs of injection before your agent acts.

Little Canary runs untrusted text through a powerless "canary" model first and watches what it does. A host integration can use the verdict to block the input or restrict downstream tool authority, depending on that host's capabilities.

Quick start · Integrations · How it works · Research · Website

The problem

A prompt injection looks like ordinary data — a web page, an email, a tool result — until your agent follows the instructions hidden inside it. Pattern-matching filters can miss attacks that do not match known patterns. By the time you notice, the agent has already acted.

The idea

Coal miners sent a canary in first. Little Canary does the same thing for agents:

1. Send the untrusted text to a canary model that has no tools, no credentials, and nothing to lose.
2. Inspect the canary's response for signs the input changed its behavior.
3. Return a verdict (PASS, FLAG, BLOCK) for the host to enforce within its available interception points.

Structural checks run alongside to catch known attack shapes. The canary can reveal attacks that no pattern check has catalogued yet.

Little Canary is developed by Hermes Labs.

Quick start (2 minutes)

Requires Python 3.9+ and a local Ollama install.

pip install little-canary
ollama pull qwen2.5:1.5b
little-canary demo --live --backend ollama --model qwen2.5:1.5b --endpoint http://127.0.0.1:11434

The demo sends one clean input and one injected input through your local canary and prints both responses, the signals it found, and the verdict for each. You'll see whether the canary followed the attack and whether Little Canary caught it.

Results depend on the canary model you run.

Screen your own input

Start the local screening service:

little-canary serve --mode block --canary-model qwen2.5:1.5b --ollama-url http://127.0.0.1:11434

Send it anything untrusted:

curl -sS http://127.0.0.1:18421/check \
  -H 'Content-Type: application/json' \
  -d '{"text":"untrusted text"}'

The service binds to loopback only and exposes GET /health. Your application reads the verdict and decides what to do before forwarding the input.

Python apps can skip HTTP and call SecurityPipeline.check() directly — see the example integrations.

The default remains qwen2.5:1.5b. Select an installed model with serve --canary-model, demo --model, or Python's SecurityPipeline(canary_model=...); the Hermes Agent plugin reads LITTLE_CANARY_MODEL. Locally exercised alternatives are qwen3.5:2b-q4_K_M, LiquidAI/lfm2.5-1.2b-instruct:q4_k_m, and gemma3:1b. Pull weights with ollama pull <tag> and check each model's license. These are selectable models, not performance guarantees; see the evaluation guidance.

Reading a verdict

Routing and coverage are separate. A fail-open setup can let a turn continue while reporting degraded coverage. Don't mistake that for a behavioral pass. Failed canary coverage can include a coverage_reason in the layer result, such as output_limit or timeout; this diagnostic does not change the verdict.

Known false-block limitation: quoted attack phrases in security reports can trigger a structural BLOCK; harmless canary acknowledgements can also trigger behavioral rules. In block mode, legitimate work can be blocked. Little Canary has no built-in pause-and-approve UI; hosts can use advisory routing and implement review where their interception point allows it. See the host capability matrix.

Put it in front of your agent

Run the local service above, then wire in the host you use. Not every host lets a plugin refuse a prompt; the host capability matrix records exactly what each one can intercept and block.

openclaw plugins install ./plugins/openclaw --force
openclaw plugins enable little-canary-openclaw
openclaw config set plugins.entries.little-canary-openclaw.hooks.allowConversationAccess true

What Little Canary is and isn't

• It's a sensing layer. It sits alongside tool policy, sandboxing, and human review — it doesn't replace them.
• The canary is powerless at the application layer, not isolated by an OS sandbox.
• A PASS means no covered signal was found in this inspection. It doesn't mean the input is safe under every circumstance.
• Detection rates depend on your model, your policy, and the attack. Benchmarks and methodology show how we measure it and where it misses.

We'd rather you know the edges than find them in production.

Research

The behavioral canarying technical note explains the powerless-model probe and why a routing decision must stay separate from whether inspection actually ran.

Contributing

Found an injection that got through? Open an issue with a safe reproducer and the observed result. For fixes or integrations, run the offline tests and Ruff, then open a pull request. The contributor guide has setup and submission details.

Benchmarks · Host capability matrix · Security policy · Releases

License

Apache-2.0. The source version lives in pyproject.toml; compare little-canary --version against GitHub Releases and PyPI for the published build.