Skip to main content

Little Canary

Sacrificial LLM instances as behavioral probes for prompt injection detection

PyPI version License: Apache 2.0 Python 3.8+ CI

What it does

  • Runs a fast structural filter (regex + decode/recheck for base64, hex, ROT13, reverse encodings)
  • Probes raw input with a small sacrificial "canary" model and checks for behavioral compromise
  • Returns either block, flag + advisory, or pass depending on mode

When to use

  • You run an LLM app or agent and want a lightweight pre-check for prompt injection
  • You can tolerate ~250ms additional latency per input
  • You want a model-agnostic layer that works with your existing stack

When not to use

  • You need formal security guarantees or audited benchmark comparability
  • You cannot accept pass-through behavior when the canary is unavailable (see Fail-open design)

Results snapshot

  • 99.0% detection on TensorTrust (400 real attacks, Claude Opus), 94.8% with 3B local model
  • 0% false positives on 40 realistic customer chatbot prompts
  • ~250ms latency per check on consumer hardware

TensorTrust benchmark — see Benchmarks and Limitations for methodology and caveats.


Table of Contents


Quick Start

# 1. Install Ollama and pull a canary model
ollama pull qwen2.5:1.5b

# 2. Install Little Canary
pip install little-canary
from little_canary import SecurityPipeline

pipeline = SecurityPipeline(canary_model="qwen2.5:1.5b", mode="full")
verdict = pipeline.check(user_input)

if not verdict.safe:
    return "Sorry, I couldn't process that request."

# Prepend advisory to your existing system prompt
system = verdict.advisory.to_system_prefix() + "\n" + your_system_prompt
response = your_llm(system=system, messages=[{"role": "user", "content": user_input}])

That's it. Your LLM, your app, your logic. The canary adds a security layer in front.

Agent Systems Quick Start

For modern agent stacks, treat Little Canary as inbound risk sensing, not your only control plane.

Recommended deployment pattern:

  1. Ingress scan all untrusted text (chat, email, web content, tool output) with pipeline.check().
  2. Block/flag using mode="full" or mode="block" depending on risk tolerance.
  3. Attach advisory (verdict.advisory.to_system_prefix()) before planner/tool decisions.
  4. Pair with outbound/runtime controls (e.g., command/domain policy monitor) for containment.

Minimal agent wrapper:

verdict = pipeline.check(untrusted_input)
if not verdict.safe:
    return {"status": "blocked", "reason": verdict.summary}

guarded_system = verdict.advisory.to_system_prefix() + "\n" + base_system_prompt
return run_agent(system=guarded_system, user_input=untrusted_input)

Little Canary is strongest when paired with runtime enforcement (outbound policy + incident logs), especially for autonomous tool-using agents.

How It Works

User Input --> Structural Filter (1ms) --> Canary Probe (250ms) --> Your LLM
                   |                            |
              Known patterns              Behavioral analysis
              (regex + encoding)          (did the canary get owned?)

Layer 1: Structural Filter (~1ms) Regex-based detection of known attack patterns, plus decode-then-recheck for base64, hex, ROT13, and reverse-encoded payloads.

Layer 2: Canary Probe (~250ms) Feeds raw input to a small sacrificial LLM (qwen2.5:1.5b by default). Temperature=0 for deterministic output. The canary's response is analyzed for signs of compromise: persona adoption, instruction compliance, system prompt leakage, refusal collapse.

Analysis Layer (pluggable)

  • Default: regex-based BehavioralAnalyzer — fast, zero dependencies
  • Experimental: LLMJudge — a second model classifies the canary's output as SAFE/UNSAFE

Advisory System Suspicious inputs that aren't hard-blocked generate a SecurityAdvisory prepended to your production LLM's system prompt, warning it about detected signals.

Why a sacrificial model?

Every existing defense classifies inputs. Little Canary observes what attacks do to a model and reads the aftermath:

  • Llama Guard evaluates content against safety categories. Little Canary detects behavioral compromise, not content safety violations.
  • Prompt Guard detects injection patterns in input text. Little Canary uses actual LLM behavioral response rather than input-side classification.
  • NeMo Guardrails uses rules and LLM calls to control dialogue flow. Little Canary works with any LLM stack, no framework required.

The canary is deliberately small and weak. It gets compromised by attacks that your production LLM might resist. That's the point — a compromised canary is a strong signal.

Deployment Modes

Mode Behavior Best For
block Hard-blocks detected attacks Customer chatbots, zero-tolerance systems
advisory Never blocks, flags for production LLM Zero-downtime systems, monitoring
full Blocks obvious attacks, flags ambiguous ones Agents, email processors, hybrid workflows

Fail-open Design

[!NOTE] If Ollama is unavailable, the pipeline passes all inputs through unscreened. This is a deliberate availability-over-security tradeoff.

How to operate safely:

  • Call pipeline.health_check() at startup to verify the canary model is reachable
  • Monitor the canary_available field in health check output
  • Alert if the canary becomes unavailable in production

Benchmark Results

Tested against an internal red-team suite of 160 adversarial prompts across 9 attack categories, plus a separate false-positive test of 40 realistic chatbot prompts.

Metric Value
TensorTrust detection rate 99.0% (400 attacks, Claude Opus as production LLM)
3B local model detection rate 94.8% (TensorTrust, 400 attacks)
Canary standalone block rate 37% (canary + structural filter alone)
False positive rate 0/40 on realistic chatbot traffic
Latency ~250ms per check

Detection by category:

Category Effective Rate Attacks
Role escalation 90% 20
Benign wrapper 70% 20
Multi-step trap 70% 20
Classic injection 65% 20
Tool trigger 65% 20
Context stuffing 50% 20
Encoding/obfuscation 40% 20
Paired obvious — 10
Paired stealthy — 10

[!NOTE] TensorTrust validated. v0.2.0 results are benchmarked against 400 real-world TensorTrust attacks. Internal red-team category breakdown above is from the original v0.1.0 test suite. See littlecanary.ai for multi-model comparison.

Integration Examples

Customer Chatbot (Block Mode)

from little_canary import SecurityPipeline

pipeline = SecurityPipeline(canary_model="qwen2.5:1.5b", mode="block")

def handle_message(user_input):
    verdict = pipeline.check(user_input)
    if not verdict.safe:
        return "I'm sorry, I couldn't process that. Could you rephrase?"
    return call_your_llm(user_input)

Email Agent (Full Mode)

from little_canary import SecurityPipeline

pipeline = SecurityPipeline(canary_model="qwen2.5:1.5b", mode="full")

def process_email(email_body, sender):
    verdict = pipeline.check(email_body)
    if not verdict.safe:
        quarantine(email_body, sender, verdict.summary)
        return
    system = verdict.advisory.to_system_prefix() + "\n" + agent_prompt
    agent.process(system=system, content=email_body)

See examples/ for complete integration code.

API Quick Reference

from little_canary import SecurityPipeline

# Initialize
pipeline = SecurityPipeline(
    canary_model="qwen2.5:1.5b",   # any Ollama model
    mode="full",                     # "block", "advisory", or "full"
    ollama_url="http://localhost:11434",
    canary_timeout=10.0,
)

# Check input
verdict = pipeline.check(user_input)
verdict.safe              # bool — is input safe to forward?
verdict.blocked_by        # str or None — "structural_filter" or "canary_probe"
verdict.advisory          # SecurityAdvisory — flagged signals
verdict.advisory.flagged  # bool — were suspicious signals detected?
verdict.advisory.to_system_prefix()  # str — prepend to your system prompt
verdict.total_latency     # float — seconds

# Health check
health = pipeline.health_check()
health["canary_available"]  # bool

Running the Benchmarks

# Red team suite (160 adversarial + 20 safe prompts, live dashboard)
cd benchmarks
python3 red_team_runner.py --canary qwen2.5:1.5b
# Dashboard at http://localhost:8899

# False positive test (40 realistic prompts)
python3 run_fp_test.py

# Full pipeline test (canary + production LLM)
python3 full_pipeline_test.py --canary qwen2.5:1.5b --production gemma3:27b --attacks-only

Project Structure

little-canary/
├── little_canary/                 # Core package (pip install .)
│   ├── __init__.py
│   ├── py.typed                   # PEP 561 type marker
│   ├── structural_filter.py       # Layer 1: regex + encoding detection
│   ├── canary.py                  # Layer 2: sacrificial LLM probe
│   ├── analyzer.py                # Behavioral analysis (regex-based)
│   ├── judge.py                   # LLM judge (experimental, replaces regex)
│   └── pipeline.py                # Orchestration + three deployment modes
├── tests/                         # Unit tests (pytest, 98%+ coverage)
├── examples/                      # Integration examples
├── benchmarks/                    # Test suites and dashboard
├── .github/                       # CI, issue templates, dependabot
├── pyproject.toml
└── requirements.txt

Troubleshooting

"Cannot connect to Ollama"

  • Ensure Ollama is running: ollama serve (or check with pgrep ollama)
  • Verify the URL: default is http://localhost:11434
  • Test connectivity: curl http://localhost:11434/api/tags

"Model not found"

  • Pull the model first: ollama pull qwen2.5:1.5b
  • The model name must match exactly (e.g., qwen2.5:1.5b, not qwen2.5)

High false positive rate

  • Use mode="full" instead of mode="block" to flag ambiguous inputs as advisories rather than hard-blocking
  • Check benchmarks/run_fp_test.py against your traffic patterns

Slow response times

  • The default qwen2.5:1.5b targets ~250ms. Set a lower canary_timeout to fail fast.
  • Use enable_structural_filter=True, enable_canary=False for structural-only mode (~1ms, no LLM required).

Limitations

  • TensorTrust validated (v0.2.0). 99.0% on 400 attacks with Claude Opus; 94.8% with 3B model. Garak and HarmBench still pending.
  • Multi-model tested (v0.2.0). Performance varies by model — see littlecanary.ai for comparison.
  • Regex-based behavioral analysis. The experimental LLMJudge is included for higher accuracy.
  • No production deployment data. All results are from controlled testing.
  • Ollama-only. No abstraction layer for other backends yet.

Roadmap

  • Benchmark against TensorTrust (99.0% detection, 400 attacks) — Garak and HarmBench still TODO
  • LLM judge to replace regex analyzer (higher accuracy)
  • Backend abstraction layer (vLLM, llama.cpp, OpenAI-compatible APIs)
  • Fine-tuned canary model (increased susceptibility = stronger signal)
  • Multi-canary ensemble for higher detection rates
  • Agent integration SDK (MCP, LangChain, CrewAI)

Contributing

See CONTRIBUTING.md for development setup and contribution guidelines.

Citation

@software{little_canary,
  author = {Bosch, Rolando},
  title = {Little Canary: Sacrificial LLM Instances as Behavioral Probes for Prompt Injection Detection},
  year = {2026},
  url = {https://github.com/roli-lpci/little-canary},
  license = {Apache-2.0}
}

License

Apache 2.0 — see LICENSE for details.

Metadata

Release files for little-canary 0.2.2

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for little-canary 0.2.2
File Size Uploaded Provenance Actions
little_canary-0.2.2.tar.gz 40.8 kB Without provenance Details for little_canary-0.2.2.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for little-canary 0.2.2
File Interpreter ABI Platform Actions
little_canary-0.2.2-py3-none-any.whl Python 3 none any Details for little_canary-0.2.2-py3-none-any.whl

Total release size: 71.5 kB

Release files / little_canary-0.2.2.tar.gz

Download URL little_canary-0.2.2.tar.gz
Size 40.8 kB
Tags Source
SHA-256 checksum
How to use checksums
c2183cd856290b29e6981148ce3f5b81841ff123aef94daedd4f2c9900a1c946
BLAKE2b-256 checksum
How to use checksums
35b4298c3135d3b8caccaf5358681960f24e61dfe032566019cc9ed1d49a8e3d
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.9.6

Release files / little_canary-0.2.2-py3-none-any.whl

Download URL little_canary-0.2.2-py3-none-any.whl
Size 30.7 kB
Tags Python 3
SHA-256 checksum
How to use checksums
5ce3a222a1d4b3d71db739b3e65f55f0b430eef79caca579ce032b9df7396d22
BLAKE2b-256 checksum
How to use checksums
f05c4890aa99a1e9d6c8a558a1615808a28ccbda87c2f5065096cf18f4530595
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.9.6

Security for little-canary 0.2.2

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page