Why Gitleaks Isn't Enough for AI Agent Security
Gitleaks is excellent at finding secrets in your code. But it doesn't understand that your MCP config just exposed those secrets to 5 connected AI tools.
TL;DR
- Gitleaks is excellent - 18k stars, 700+ patterns, industry standard for secret scanning
- But it treats agent configs as regular files. It doesn't understand agent topology.
- Agent-aware scanning maps which tools can reach which credentials - context changes severity
- Keep Gitleaks for your code. Add Firmis for your agent stack.
Gitleaks Is a Great Tool
Let's start here: if you're not running Gitleaks, start now. 18k+ GitHub stars, 700+ detection patterns, battle-tested across millions of repos. It's the industry standard for finding secrets in your codebase.
This post isn't about replacing Gitleaks. It's about what it was never designed to do.
The Gap: Agent-Aware Detection
Gitleaks finds secrets. Firmis understands what can reach them.
Gitleaks finds
Found the secret. Job done.
Firmis finds
Found the secret + what can reach it + what that means.
Finding the secret is step one. Knowing what can reach it is the part that matters.
What 'Agent-Aware' Means
- Maps which tools can read which credentials
- Understands that ~/.aws/credentials in an MCP config means ALL connected tools can access it
- Detects that file access + network access = potential exfiltration vector
- Traces the agent → tool → credential → network topology
What Gitleaks Misses in Agent Stacks
Tool poisoning patterns
Known-malicious tools
Permission over-grants
MCP server topology
Prompt injection payloads
Cross-platform threats
The Recommendation
Use Both
- →Keep Gitleaks for your codebase - it's excellent at what it does
- →Add Firmis for your agent stack - agent-aware scanning that understands topology
- →They're complementary: Gitleaks protects your code, Firmis protects your agent infrastructure
Generic secret scanning finds what's exposed. Agent-aware scanning tells you who can exploit it.
Try It Now
Find out if your agent stack is safe
One command. 30 seconds. Free.
Fix and Monitor included with Pro
View pricing