Stop prompt injection before it reaches your production agent
Repellox blocks malicious inputs, jailbreaks, and unauthorized tool calls in real time. Model providers do not include this guardrail. You need it.
| Timestamp | Status | Threat Class | Risk Score |
|---|---|---|---|
| 14:22:07.114 | BLOCKED | prompt_injection:system_override | 0.97 |
| 14:22:08.331 | ALLOWED | clean | 0.02 |
| 14:22:09.882 | BLOCKED | jailbreak:role_confusion | 0.91 |
| 14:22:11.004 | BLOCKED | tool_abuse:unauthorized_exfil | 0.88 |
| 14:22:12.750 | ALLOWED | clean | 0.04 |
Model providers ship intelligence, not security
Each AI agent you deploy is a live attack surface. Each model API brings three failure modes.
Prompt Injection
Malicious instructions in user inputs or retrieved documents can override your agent system prompt and completely redirect its behavior.
Jailbreaks
Adversarial wording can make models abandon safety guidelines, generate harmful outputs, or expose confidential system context.
Unauthorized Tool Calls
Agents with tool access may be coerced into calling APIs, reading files, or exfiltrating data beyond the intended scope.
Inline interception, no refactoring
Repellox sits between your application and the model API. One SDK call, no pipeline rewrites.
Wrap in one line
Swap your model client for the Repellox SDK. Your prompts, pipeline, and infrastructure stay unchanged.
Every request is scored
Repellox evaluates each input against 40+ threat signatures, semantic risk models, and your custom policy rules in under 2ms.
Threats are stopped, not just logged
High-risk requests stop before the model. Clean requests pass through untouched. Review the threat log from your dashboard.
Built for the production attack surface
Each layer addresses a distinct AI agent threat.
Injection Classifier
Trained on thousands of injection samples, the classifier detects indirect injections in retrieved documents and direct user inputs.
Policy Engine
Define what your agent may do. Block specific tools, limit output topics, and enforce role boundaries through a declarative YAML config.
Threat Dashboard
Every blocked request includes full context: original input, threat class, risk score, and decision trace. Audit-ready from the start.
Low-Latency Runtime
Repellox adds under 2ms of latency. Deployed on Cloudflare's edge network, the interceptor co-locates with your inference endpoint globally.
SDK-First Integration
Python and TypeScript SDKs support LangChain, LlamaIndex, and direct OpenAI / Anthropic clients first-class. Run your agent anywhere.
Zero Data Retention
Repellox scores and discards payloads. By default, nothing is stored on Repellox servers. SOC 2 controls framework, GDPR-ready handling.
Teams shipping AI agents ship with Repellox
We caught three injection attempts in our RAG pipeline in week one. Our security team finally stopped asking us to delay the AI rollout.
We imported one SDK and had an intercept log. No infrastructure changes or prompt-pipeline refactoring. It took an afternoon.
Injection detection accuracy
Measured on a held-out red-team dataset with 8,400 adversarial inputs across 12 threat classes.
Simple pricing, no per-token overhead
Pay for inspected requests, not processed tokens.
Your AI agent is in production. Is its security?
Protect your agents in 15 minutes. Free tier, no credit card required.
Questions? [email protected]