GuardAgent: SlackBoard-Native DevSecOps AI

GuardAgent is an AI DevSecOps agent native to Slack. It uses Gemini to scan live GitLab commits, maps threats to CWE IDs, and mentors developers with instant remediation code right in their slack chat

Comment

💡 Inspiration

I have always wanted to see how cybersecurity could seamlessly interface with artificial intelligence. If teams are working remotely, they should be able to see what changes are happening in real-time. Slack is the central nervous system for most modern technology departments. The goal is to build a ChatOps tool that doesn't just block broken code, but actively mentors the engineering team.

⚙️ How I built it

GuardAgent is a fully event-driven architecture designed for speed and precision.

  • Backend Infrastructure: Deployed as a serverless Python FastAPI application on Google Cloud Run for instant auto-scaling.
  • Model Context Protocol (MCP): I completely architected the backend as an MCP server, implementing Server-Sent Events (SSE) transport to standardize how the AI interacts with internal developer tools.
  • The AI Engine: We utilized the google-genai SDK to power the core logic with gemini-2.5-flash. The AI reads the raw code diffs and maps threats directly to MITRE CWE IDs.
  • Threat Scoring: To prioritize alerts, the AI calculates a threat severity score. The overall severity score is calculated based on the base CWE weight and the AI's confidence interval for the detected vulnerabilities:

$$ S = \sum_{i=1}^{n} (W_i \times C_i) $$

  • Integrations: Connected via GitLab Webhooks to detect live code pushes, and authenticated via the GitLab API to pull exact commit diffs.
  • ChatOps Interface: Built using the Slack Bolt API, leveraging rich Block Kit UI elements (including dynamic scanning visuals) to deliver interactive, highly readable alerts.
  • Security & Secrets: Enforced a zero-trust architecture by vaulting all API keys, Slack tokens, and webhook signatures inside Google Secret Manager, alongside custom dependency injection to secure the MCP endpoints.

🚧 Challenges I ran into

  • Model Availability: I initially engineered the platform for the brand-new gemini-3.5-flash model but hit regional data center availability limits (404 NOT_FOUND) on Google Cloud. I quickly re-engineered the architecture to utilize the highly stable gemini-2.5-flash without losing any analysis capabilities.
  • Contextual Memory: Initially, the bot treated every Slack mention as a blank slate. I engineered a custom state manager mapping Slack Thread IDs (thread_ts) to persistent Gemini Chat Sessions, allowing the AI to remember the code context across multiple follow-up questions.
  • Protocol Security: Implementing the MCP server required strict security controls to ensure external access to the AI toolset remained completely locked down via a custom verify_mcp_key dependency.
  • Translating Syntax: Bridging standard Markdown and raw code outputs into Slack's proprietary formatting and Block Kit JSON required meticulous regex parsing to ensure the UI remained pristine.

🏆 Accomplishments that I'm proud of

  • Authentic End-to-End Flow: GuardAgent authentically reaches into live GitLab repositories, parses real commits, and streams genuine diffs to the LLM. There is zero mocked data in the backend pipeline.
  • MCP Integration: Successfully upgrading the bot to an official Model Context Protocol server, proving that the tool can act as a secure, standardized bridge between Slack, GitLab, and Gemini.
  • Automated CWE Compliance: The bot doesn't just find bugs; it surgically identifies them and maps them to global MITRE CWE standards for instant compliance reporting.
  • Interactive Mentorship: I successfully evolved the tool from a one-shot scanner into an interactive security mentor that remembers thread context and writes remediation patches on demand.

📚 What I learned

I gained deep experience designing strict system prompts to prevent LLM hallucinations, ensuring the AI only flags genuine threats rather than standard boilerplate code. I also mastered the Model Context Protocol standard, learning the complexities of bridging event-driven webhooks securely across GitLab, Google Cloud, and Slack.

🚀 What's next for GuardAgent

The next phase involves expanding webhook support to GitHub Advanced Security and giving GuardAgent the capability to automatically open functional Merge Requests with its AI-generated remediation code.

Built With

Share this project:

Updates