gjallarhorn-hq-sdk 1.0.1

Official Python SDK for the Gjallarhorn prompt-injection detection API

gjallarhorn-hq-sdk

Official Python SDK for the Gjallarhorn prompt-injection detection API.

Installation

pip install gjallarhorn-hq-sdk

Requirements: Python 3.9+, httpx.

Quickstart

import os
from gjallarhorn_hq_sdk import GjallarhornClient

client = GjallarhornClient(api_key=os.environ['GJALLARHORN_API_KEY'])

result = client.scan(input=user_input)
if result.risk_level in ('critical', 'high'):
    raise ValueError('Injection detected — request blocked')
# safe to send to LLM

Async usage

import asyncio
import os
from gjallarhorn_hq_sdk import AsyncGjallarhornClient

async def main():
    async with AsyncGjallarhornClient(api_key=os.environ['GJALLARHORN_API_KEY']) as client:
        result = await client.scan(input='some user input')
        print(result.risk_level, result.detection_layers)

asyncio.run(main())

Methods

Method	Description
scan(content, *, use_classifier?, context?)	Scan text for prompt injection
scan_multimodal(file, mime_type, filename?)	Scan PDF/image/QR code
get_quota()	Get credit usage and quota
check_canary(llm_response)	L2 output integrity check (opt-in)

Both GjallarhornClient (sync) and AsyncGjallarhornClient (async) expose the same methods.

Scan a PDF

from pathlib import Path

result = client.scan_multimodal(Path('user-upload.pdf'), 'application/pdf')
print(result.risk_level, result.billed_pages)

L2 Agentic Integration

check_canary() checks whether your LLM's response still contains the Gjallarhorn canary token. Absence signals a possible system-prompt override. L2 is opt-in — it only covers agents explicitly registered with Gjallarhorn; LLM calls made outside the integration are not monitored.

Privacy: L2 canary checks are stateless and content-free. Gjallarhorn receives the LLM output, performs a token presence check, and returns a boolean result. No output content, user data, or conversation text is retained server-side.

Full L2 integration example

import os
from gjallarhorn_hq_sdk import GjallarhornClient
from gjallarhorn_hq_sdk import AuthError, ServiceUnavailableError

client = GjallarhornClient(api_key=os.environ['GJALLARHORN_API_KEY'])

# Step 1: Register agent once at startup — store canary_token in your secrets.
# (Registration is a one-time REST call to POST /v1/agents/register.)
CANARY_TOKEN = os.environ['AGENT_CANARY_TOKEN']

# Step 2: Inject canary into every system prompt.
def build_system_prompt(base_prompt: str) -> str:
    return f"{base_prompt}\n{CANARY_TOKEN}"

# Step 3: After every LLM call, check output integrity.
def check_agent_output(llm_output: str) -> None:
    try:
        check = client.check_canary(llm_output)
    except AuthError:
        raise RuntimeError('Canary check: invalid API key')
    except ServiceUnavailableError:
        # Canary service down — decide whether to fail-open or fail-closed
        import warnings
        warnings.warn('Canary check unavailable — proceeding with caution')
        return

    if check.alert_code == 'RAGNARÖK':
        # Canary absent: model did not echo the token it should always produce.
        # Possible cause: system prompt was overridden by an injection attack.
        raise RuntimeError(
            f'[RAGNARÖK] Agent output integrity check failed '
            f'for agent {check.agent_id} at {check.checked_at}'
        )

# Usage in your agent loop:
system_prompt = build_system_prompt('You are a helpful customer support agent.')
llm_output = call_your_llm(system_prompt, user_message)
check_agent_output(llm_output)
# Safe to proceed

Error handling

from gjallarhorn_hq_sdk import (
    AuthError, QuotaExceededError, RateLimitError,
    ExtractionFailedError, ServiceUnavailableError,
)

try:
    result = client.scan(input=user_input)
except AuthError:
    print('Invalid API key')
except QuotaExceededError:
    print('Monthly quota hit')
except RateLimitError as e:
    print(f'Rate limited — retry after {e.retry_after}s')
except ExtractionFailedError:
    print('Could not extract file content')

The SDK retries automatically on rate-limit (respecting retry_after) and 503 errors (exponential backoff, max 3 retries).

License

MIT — see LICENSE.

Links

• API Reference
• Documentation
• Issues