How many secrets are exposed in MCP config files?

GitGuardian's 2026 report found 24,008 unique secrets exposed in public MCP configuration files on GitHub, including 2,100+ confirmed valid credentials.

What solutions exist for agent credential management?

Agent Vault (1,661 stars), OneCLI (2,369 stars), and Zerobox (649 stars) act as credential proxies. They intercept tool calls and inject secrets at the transport layer so the agent sees only a JWT, not raw API keys.

What product could be built here?

A universal credential vault that embeds directly into the MCP protocol as middleware: agents register with a single identity, the vault vends scoped time-limited tokens per tool call, and every key use is audit-logged.

← Back to dashboard

clawsmith.com/signal/mcp-config-credential-sprawl-93pct-unscoped-api-keys

⚠ IssueUnderservedLive

93% of AI agent projects expose unscoped API keys in MCP config files

Q: Why do AI agents leak API keys through MCP?

MCP quickstart guides recommend putting API keys directly in claude_desktop_config.json and .cursor/settings.json. One prompt injection can cause an agent to read and exfiltrate those keys.

Q: What percentage of AI agent projects use unscoped API keys?

93% of AI agent projects studied use unscoped API keys as their only auth mechanism, 0% have per-agent cryptographic identity, and 100% lack per-agent revocation capability.

GitGuardian's 2026 report found 24,008 unique secrets in public MCP configuration files on GitHub. A study of 30 AI agent projects found 93% use unscoped API keys as their only auth mechanism, 0% have per-agent cryptographic identity, and 100% lack per-agent revocation. MCP quickstart guides normalize putting API keys directly in claude_desktop_config.json and .cursor/settings.json. One prompt injection leaks everything.

Product Idea from this Signal

A web app that intercepts MCP tool calls and injects scoped, time-limited credentials so AI agents never hold raw API keys

2.5k ▲

93% of AI agent projects paste raw API keys directly into MCP config files, creating a credential leak surface that scales with every tool added. Developers have no native way to scope, rotate, or audit which keys an agent actually used. This web app sits as a lightweight proxy between the MCP client and the upstream APIs, injecting short-lived credentials at call time and revoking them automatically, so a compromised agent config exposes nothing.

ai-agentsmcpsecrets-managementcredential-securitydeveloper-tools

Competitive102 leadsView Opportunity →