Cerberus

An intelligent API security proxy that shields users and AI services from bad actors, serving as an invisible layer of defense against exploitation and digital vulnerabilities.

Comment

Inspiration

Indonesia is rapidly improving and integrating AI into its many systems like in banking systems, customer services, chatbots and more. These advancements bring severe concerns for the general public, particularly among communities navigating the digital ecosystem. For these citizens, an official bank or government chatbot represents absolute trust. However, without proper protection, these AI systems can be heavily misused by bad actors for fraud, identity theft, and generating malicious content. We were inspired to build a system that protects this rapid automation push, ensuring that technological progress does not come at the cost of public safety

What it does

Cerberus is a zero-trust security gateway for LLM traffic that sits seamlessly between your application and any OpenAI-compatible model endpoint. It inspects every request and response in both directions, completely neutralizing threats before they can reach the model or your users.

  • One-Line Integration: Developers don't need to alter their application code. By pointing their OpenAI client's base URL to the Cerberus proxy endpoint, they gain immediate security. It natively speaks the OpenAI Chat Completions schema and routes traffic to any configured backend whether it is OpenRouter, vLLM, Ollama, or a custom model.
  • Inbound Defense: Every incoming request is deeply evaluated, decoded, and stripped of malicious payloads. It aggressively sanitizes traffic to block direct prompt injections, jailbreaks, obfuscated strings, and indirect injections hidden in external tool outputs or retrieved documents.
  • Outbound Defense: Model outputs are monitored via full buffered scanning or near-live streaming. Cerberus screens for leaked corporate secrets, PII (like banking data or ID numbers), and system-prompt echoes, cutting off the connection before the compromised information reaches the client. Server-Side Key Isolation: To prevent upstream security breaches, provider keys stay strictly server-side. Clients authenticate to Cerberus using unique, low-privilege tokens; our gateway swaps in the real upstream credential right before forwarding the payload, keeping the provider keys safe within your infrastructure.

How we built it

We engineered Cerberus from the ground up to handle massive concurrency with very low latency overhead, ensuring that adding a security layer wouldn't degrade the conversational user experience.

  • The Gateway Engine (Go): The core zero-trust proxy was built entirely in Go to ensure maximum throughput fast execution speeds. Go’s concurrent routing mechanics allowed us to parse incoming streams, swap authorization tokens server-side, and process mathematical heuristic rule checks simultaneously.
  • The Control Plane (Vue & TypeScript): The centralized web dashboard was developed using TypeScript and Vue.js. This control plane lets teams configure their upstream routing, dynamically issue or revoke client keys, and toggle specific safety detectors live.
  • Visual Testing Environment: We leveraged Tailwind CSS to build a modern interface featuring a built-in Test Zone. This interactive module lets developers safely fire adversarial text attacks at their endpoints and watch the structured security verdicts update in real time

Challenges we ran into

  • We realized that traditional firewalls are not reliable enough, the challenge is we need to find an optimal solution into solving the problem.
  • Initially, team members weren't communicating efficiently, due to different work shifts. After assembling, we managed to finish the project according to our plan

Accomplishments that we're proud of

  • A Fully Working Zero-Trust Architecture: We built a highly concurrent proxy layer that successfully blocks malicious intent from attackers, masks private user data, and stops prompt manipulation out of the box.
  • The Developer Control Plane: We successfully engineered a visually striking web dashboard complete with an automated Test Zone, giving teams an instant sandbox to visualize and verify their security posture.
  • Structured Verdict Output: Every single request traveling through Cerberus exits with an immutable, structured security log, meaning developers always know exactly what text was allowed, flagged, or blocked and the precise heuristic rule why.

What we learned

  • Teamwork between team members is important, proper teamwork and communication skills increase productivity and efficiency.
  • Architecturally, we learned that a standalone traditional proxy isn't enough, true security requires decoupling your critical upstream model credentials from the client side completely.

What's next for Cerberus

We intend to mature Cerberus into a highly accessible and reliable security Software as a Service (SaaS) platform. We are planning to expand our inbound parsing rules to deeply support localized compliance frameworks, automatically generating audit-ready logs that comply with regional OJK financial regulations and national BSSN security mandates out of the box. We will also expand the control plane's Test Zone to include pre-configured simulation profiles, allowing developers to test their applications against complex regional cyberthreat scripts with a single click.

Built With

Share this project:

Updates