Calus: The AI Gateway for Prompt Injection Detection

Research Context and Background

Calus is a security tool designed to protect applications that use artificial intelligence. It acts as a middle layer between your software and AI providers like OpenAI or Anthropic. Its main job is to scan every message sent to an AI model to find dangerous attempts to trick the system. These tricks, known as prompt injections or jailbreaks, try to make the AI ignore its safety rules or leak private information. Calus spots these threats before they reach the AI, ensuring your application stays safe without needing complex changes to your existing code.

Benefits

Calus offers several key advantages for developers and security teams. First, it is detection-only, meaning it identifies threats but does not block or change the traffic. This allows teams to see exactly what is happening without disrupting their workflow. Second, it requires zero code changes. Users only need to set one environment variable to start protecting their applications. Third, it is very fast. Scanning takes less than 15 milliseconds per request, so it does not slow down the application. Finally, it is secure and private. The tool never stores the API keys used to connect to AI providers, keeping sensitive credentials safe.

Use Cases

Calus is useful for any application that interacts with large language models. It is particularly helpful for chatbots, customer service agents, and tools that browse the web or execute code on behalf of users. For example, if a chatbot reads a document that contains hidden instructions to steal data, Calus can spot that attempt. It also helps when an AI tool is asked to run code, as it can catch dangerous commands before they are executed. Developers can use it to monitor their systems in real time through a live dashboard that shows flags and confidence scores for every request.

Pricing

Calus is available as open-source software under the MIT license. This means it is free to use, modify, and distribute. There are no hidden fees or account requirements to start using the tool. Developers can run it locally or on their own servers without paying for a subscription.

Vibes

As an open-source project, Calus has not yet gathered a large number of public reviews or testimonials. However, its design focuses on high accuracy based on academic benchmarks. Tests show it achieves 100% recall on specific types of prompt injections and jailbreaks. It also maintains a very low false positive rate, meaning it rarely mistakes normal user messages for threats. This balance of high detection and low noise makes it a promising tool for securing AI applications.

Additional Information

Calus is built to work with many popular AI tools and frameworks. It supports SDKs from OpenAI, Anthropic, Groq, and LangChain, as well as any provider supported by LiteLLM. The tool uses a three-step scanning process to ensure speed and accuracy. It starts with fast rules to catch known patterns, then uses similarity matching for variations, and finally uses an optional semantic model for new types of attacks. All of this runs without the need for expensive GPU hardware.