What is coding agent security?

Coding agent security is the discovery and control of AI coding assistants and agentic dev tools such as Cursor, Claude Code, GitHub Copilot, and Windsurf, plus the local MCP servers and skills they use. It covers what those tools can read, run, and send, on developer machines and in CI.

Why are coding agents a security risk?

Coding agents run with a developer's privileges and can read source and secrets, execute shell commands, install packages, and call MCP servers and external APIs. Personal API keys, auto-approve settings, tool poisoning, and malicious skills can turn a helpful assistant into an exfiltration or supply-chain path that endpoint and code-scanning tools do not see.

How does Wingback secure Cursor, Claude Code, and Copilot?

A browser sensor and native macOS/Windows agents discover which coding agents, MCP servers, and skills are present, flag risky posture against the OWASP MCP Top 10, and enforce policy inline. Hooks inside coding agents (for example Claude Code PreToolUse) can warn, mask, or block a tool call before it runs, even offline.

Can Wingback detect shadow AI coding tools?

Yes. Wingback finds unsanctioned coding assistants, personal-account usage, local Ollama or LM Studio instances, and MCP servers on managed devices, and attributes them to a single device identity across the browser and native sensors.

Security for AI coding agents · Wingback Security

Cursor, Claude Code, GitHub Copilot, and Windsurf now sit on most engineering laptops. They read your source and secrets, run commands, and call MCP servers. Wingback discovers every one and controls what it is allowed to do, on the device and before a risky action runs.

The problem

A coding agent runs with the developer’s own privileges. It can open any file, execute a shell command, install a dependency, and reach an external API or MCP server. That is what makes it useful, and also what makes it dangerous. Personal API keys move company code to unsanctioned providers, auto-approve settings let an agent act without review, a poisoned tool or a malicious skill can turn the assistant into an exfiltration path, and none of it shows up in endpoint EDR or static code scanning. The tools live below the layer those controls watch.

How Wingback secures coding agents

Discover on every device. A browser sensor (Chrome and Edge) plus native macOS and Windows agents sweep for coding assistants, MCP servers, skills, and local models every few minutes, and merge into one device identity.
Score the posture. 20+ checks mapped to the OWASP MCP Top 10 flag hardcoded credentials, cleartext transport, tool poisoning, auto-approve configs, and homoglyph server names, backed by a live feed of malicious-skill indicators.
Enforce at the edge. An embedded policy engine and hooks inside coding agents (for example Claude Code PreToolUse) can warn, mask, or block a tool call before it executes, and hold even when the device is offline.
Catch shadow AI. Unsanctioned tools, personal-account usage, and data leaving to unapproved providers are surfaced and attributed.

Coverage

Coding agents: Cursor, Claude Code, GitHub Copilot, Windsurf, and generic AI web apps.
Local MCP servers (stdio), agent skills, and local models (Ollama, LM Studio).
Rollout via MDM (Intune, Jamf, GPO) with a 60-second heartbeat and one device identity across sensors.

Secure your coding agents