Classify threats before they reach AI tasks
The agent classifies injection, jailbreak, harmful intent, policy evasion, sensitive data, privacy, obfuscation, and tool manipulation threats, so nothing unsafe ever reaches your operations.
Safeguard AI-powered processes from unsafe prompts
User-supplied prompts can expose AI workflows to injection attacks, jailbreaks, and policy violations; risks that grow as AI tasks become more central to business processes. The AI Firewall Agent evaluates each prompt before it reaches an AI task, classifies the threat type, and sanitizes unsafe content so the workflow can continue without starting over. When the model's confidence falls below your threshold, the agent refines its analysis and retries automatically. The result is a structured JSON decision your BPMN process can branch on immediately. Add it in front of any AI task that processes user input and configure its behavior through process variables.
Features and Benefits
Retry automatically on low-confidence decisions
Confidence-based retry for more reliable decisions. When the model's confidence falls below your threshold, the agent refines its prompt and automatically retries. You control how many attempts are made and what confidence level is acceptable.
Sanitize unsafe prompts and enforce size limits
When a prompt is flagged as warn or block, the agent returns a cleaned version with unsafe directives removed. Legitimate intent is preserved so the process can continue without having to start over.
Connect to any supported AI provider
Works with any LLM supported by Camunda: OpenAI, Azure OpenAI, Ollama, and beyond.
Validate against real-world threat scenarios
Includes an automated test suite that validates prompt classification against a real LLM, covering block, warn, and allow scenarios out of the box.
Details
- Marketplace release date -
- Last Github commit -
-
Associated Product Group Categories:
- Patterns
-
Version Compatibility:
-
Used resources:
522488