AI Agent Security Guide
The definitive guide to securing AI coding assistants. Learn the threats, understand the attack surface, and scan your setup in 30 seconds.
Free scan. No sign-up required.
What Are the Main AI Agent Security Threats?
AI coding assistants operate with elevated privileges: file access, network calls, shell execution. Each capability expands the blast radius of an attack, meaning what an agent can actually reach and affect during a session: credentials, files, network endpoints, and other connected agents. Firmis maps this blast radius and monitors it at runtime.
Credential Harvesting
AI agents with file system access can be instructed to read ~/.aws/credentials, .env files, and SSH keys.
Data Exfiltration
Malicious tools send conversation history, source code, and internal documents to external servers.
Prompt Injection
Hidden instructions in documents, web pages, or tool outputs override the agent behavior.
Tool Poisoning
MCP server tool descriptions contain embedded instructions that hijack agent actions.
Privilege Escalation
Agents granted broad permissions can be manipulated into executing actions beyond their intended scope.
Supply Chain Attacks
Compromised dependencies or MCP servers introduce backdoors that activate during agent sessions.
How to Secure Each AI Agent Platform
Each AI platform has unique attack surfaces. Firmis covers all of them.
Claude Code
Scan Claude Desktop extensions, MCP server configs, and rules files for credential exposure and prompt injection.
Cursor
Detect malicious Cursor extensions and MCP integrations that exfiltrate code or inject instructions.
MCP Servers
Deep analysis of Model Context Protocol server configurations for tool poisoning and path traversal.
Codex / OpenAI
Scan Codex plugin definitions and OpenAI tool configurations for unauthorized data access patterns.
AI Agent Security Tools Compared
Most security tools were built before agentic AI existed. Here is how the current landscape compares for teams securing AI agent stacks.
| Tool | Platforms | Open Source | Detection Rules | Auto-Fix | Runtime Monitor |
|---|---|---|---|---|---|
| Firmis | Any AI agent platform | Yes (Apache-2.0) | Full-stack, 89% precision | Yes | Yes |
| mcp-scan (now Snyk) | MCP only | Yes (acquired by Snyk) | MCP config patterns | No | No |
| Cisco Skill Scanner | MCP skills | No | Skill behavior patterns | No | No |
| AgentAudit | Agent repos | Partial | Threat patterns | No | No |
| Snyk Agent Scan | Skills only | No | Limited | No | No |
| Lasso Security | Enterprise agents | No | Behavioral analytics | No | Yes |
| Pillar Security | Coding agents | No | Discovery + audit | No | Yes |
Why Open Source?
Security tools should not be black boxes. When you run a scanner against your agent configuration, you deserve to see exactly what rules it applies and why. The Firmis scanner is Apache-2.0 licensed and publicly auditable on GitHub.
Open source also means the detection rules improve continuously. Researchers, security teams, and platform builders contribute patterns as new threats emerge. The rule set that ships with Firmis today reflects findings from production AI agent deployments across many platforms.
The scanner is free forever. Pro features like auto-fix, active monitoring, and deep AI analysis are where the platform scales for teams.
View firmis-scanner on GitHubHow to Scan Your AI Agent Setup
Firmis runs without installation via npx. Zero config for most setups.
# Initialize and scan all platforms (recommended)
# Use in CI/CD pipelines
# Generate HTML report
Deep-Dive Security Guides
Frequently Asked Questions
Do I need a license key to scan my AI agent setup?
No. The scan command is free and works without a license key via npx firmis-cli init. It covers any AI agent platform and applies all detection rules at no cost.
What platforms does Firmis scan?
Firmis scans any AI agent platform: Claude Code, MCP servers, Codex, Cursor, CrewAI, AutoGPT, OpenClaw, Nanobot, and Supabase AI. All platforms are covered in a single scan command.
How long does a scan take?
A full scan across all installed platforms typically completes in under 30 seconds. No code is executed. Firmis performs static analysis on configuration files and tool definitions.
Scan Your AI Agent Setup Now
One command. Any AI agent platform. Know your risk in 30 seconds.