Inline security between your app and model API
Repellox checks every request between your application and the LLM API for injection signatures, jailbreak patterns, and tool authorization violations in under 2ms.
Five layers, one wrapper call
Each layer addresses a different AI agent threat. All run in parallel before the request reaches the model.
Semantic Injection Detection
Trained with real adversarial samples from 12 threat classes. Detects indirect injections in retrieved documents, not only obvious direct user inputs.
Indirect Injection in Tool Outputs
Search results and fetched documents are inspected when your agent calls a tool. RAG results are the most commonly missed surface for indirect injection.
Tool Call Authorization Rules
Set the tools your agent may call. Attempts outside the allowlist are blocked before the tool executes.
Output Inspection
Before reaching the caller, model outputs are checked for exfiltration patterns and policy violations. This blocks leakage through the response path.
Streaming Support
Streaming is supported. Chunk-level checks keep latency low while inspection stays active across the full response stream.
Repellox in your pipeline
One integration point. Leave your prompts, model config, and infrastructure unchanged.
Works with your agent framework
No lock-in. Repellox wraps your existing executor, so you can switch frameworks without changing security configuration.
Interceptor catches threats
Forty-plus signatures span four attack families, with continuous updates from active red-team research.
-
Direct Prompt Injection
User messages can contain instructions that override or hijack the system prompt, including role confusion attacks.
-
Indirect Injection via External Data
Retrieved web pages, uploaded documents, database records, and search results can conceal adversarial payloads returned to the agent.
-
Jailbreak Patterns
Phrasing designed to evade safety guidelines includes DAN variants, role-play exploits, and encoding-based obfuscation.
-
Unauthorized Tool Calls
Calls outside the declared authorization policy include tool spoofing and parameter injection attempts that expand scope.
-
Data Exfiltration via Output
Output scanning blocks attempts to expose sensitive context window contents through model responses before they reach the caller.
-
Context Window Poisoning
Multi-step attacks can corrupt early context and steer downstream agent decisions through a long-running workflow or conversation thread.
Protect your agents today
Free tier includes core injection detection. No credit card required.