---
title: "Discussion on \"Your AI Agent Got Poisoned Through a Tool Description. Here's Why Only the Kernel Saw It Coming.\" | Hashnode"
description: "Discussion on \"Your AI Agent Got Poisoned Through a Tool Description. Here's Why Only the Kernel Saw It Coming.\". MCP tool poisoning hides attacker instructions in tool metadata — invisible to you, fully visible to the model. We built kernel-level eBPF monitoring that catches it where app-layer defenses can't."
url: "https://hashnode.com/posts/your-ai-agent-got-poisoned-through-a-tool-description-here-s-why-only-the-kernel-saw-it-coming/69e69479c9501dd01034b697"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Your AI Agent Got Poisoned Through a Tool Description. Here's Why Only the Kernel Saw It Coming.](https://correlic.hashnode.dev/your-ai-agent-got-poisoned-through-a-tool-description-here-s-why-only-the-kernel-saw-it-coming?utm_source=hashnode&utm_medium=feed)

MCP tool poisoning hides attacker instructions in tool metadata — invisible to you, fully visible to the model. We built kernel-level eBPF monitoring that catches it where app-layer defenses can't.

Apr 20 · 13 min read · correlic.hashnode.dev  
Read the article

-   [#ai-security](https://hashnode.com/tag/ai-security)
-   [#cybersecurity](https://hashnode.com/tag/cybersecurity)
-   [#ebpf](https://hashnode.com/tag/ebpf)
-   [#devsecops](https://hashnode.com/tag/devsecops)
-   [#ai-agents](https://hashnode.com/tag/ai-agents)

## Written by

[Correlic](https://hashnode.com/@correlic)

@correlic

  
[View profile](https://hashnode.com/@correlic)

Kernel-level security observability for AI coding agents

AI coding agents run on your machine with broad access to files, network, and system calls — but most developers have zero visibility into what they actually do. Correlic is building kernel-level security observability to change that. Using eBPF to trace every syscall, we detect anomalous behavior, correlate threat patterns, and give developers a real-time audit trail of AI agent activity. This blog documents the technical journey of building Correlic from the ground up — the engineering challenges, the failed approaches, and the breakthroughs in making AI agents transparent and accountable.

## Responses

Markdown supported · ⌘↵ to post

No responses yet.