Malornot

• 

  Red neon shield with a lock in the center on a black background.

Inspiration

Enterprises adopting LLMs often rely on a patchwork of AI tools such as moderation, compliance, and threat detection, but these tools do not talk to each other. There is no intelligent layer that knows when to use which tool. That gap inspired us to build an agent that can orchestrate them.

What we built

We created an agent that sits in front of an LLM and dynamically decides which protection to fire based on the input. Instead of a rigid pipeline or a single filter, it understands context and routes queries through the right tools, blocking malicious prompts, applying moderation, or enforcing compliance. The result is safer, more reliable LLMs without slowing them down. Due to time constraints in this hackathon, we focused on a single-model agent. This is not a multimodal system, though future iterations could support multimodal inputs. We are also aware that there are many ways to insert malicious content into a system, but for this demo we focused on the three most common vectors in enterprise use cases.

What we learned

The problem is not a shortage of tools, it is the lack of a brain that can coordinate them. With MCP and supporting APIs, we showed that even in a hackathon weekend you can create this missing orchestration layer.

Challenges

The biggest challenge was scope. With so many possible protections, we focused on delivering a minimal but functional demo that proves adaptive orchestration works in practice.

Why it matters

Enterprises do not need another standalone tool, they need a way to make the ones they already trust work together. Our agent provides that reusable security and governance layer, giving companies confidence to scale LLM adoption responsibly today.

Continuation of the project

Looking ahead, we see two key directions:

1. Multimodal support: Extending the orchestration layer beyond text to include images, audio, and video, creating a more versatile security and compliance agent.
2. Observability and dashboards: Giving enterprises full visibility into which tools were triggered, why decisions were made, and how outcomes were enforced, building trust through transparency.

Built With

• anthropic
• dify
• llamaindex
• minimax
• qodo