Agent security
AI agents have access that traditional applications never needed, and that access creates security challenges.
What agents can access
| Access type | Examples | Risk |
|---|---|---|
| Prompts and attachments | Contracts, HR spreadsheets, customer exports pasted or attached in claude.ai and Cowork | Sensitive data exposure |
| Terminal | Bash commands, package installation, system config | Arbitrary code execution |
| File system | Source code, .env files, SSH keys, configs | Credential exposure |
| MCP tools | CRM, email, databases, APIs, cloud services | Data exfiltration, unintended writes |
| Network | HTTP requests, database connections | Unauthorized access |
Why security tools can't see what AI agents do
You can't govern what you can't see. Traditional security tools don't capture agent behavior because:
- Agent actions span multiple systems and interfaces
- Decision-making happens inside the AI model
- Tool usage changes dynamically based on context
- Most monitoring tools were built for human behavior
AI agent security risks
Sensitive data in prompts
People attach and paste the material they're working on, so prompts in claude.ai and Claude Cowork carry customer lists, contract terms, salary data, and financial forecasts. Claude inference hooks send each governed prompt to Agent Monitor, where rules can block it before it reaches the model.
Actions through connected tools
An agent with MCP tools can send email, update CRM records, or export data on a user's behalf. The MCP gateway controls which tools each team gets, and Agent Monitor records every call.
Sensitive file access (coding agents)
Agents can read files you'd rather they didn't:
- Credentials:
.envfiles, SSH keys, API tokens, database passwords - Production config: Connection strings, deployment scripts, IaC files
- Business data: Proprietary code, customer data, financial information
Command execution (coding agents)
With bash access, coding agents can:
- Read secrets from files or run
kubectl get secrets - Deploy to production without approval workflows
- Install packages that introduce vulnerabilities
- Call cloud APIs with available credentials
Prototyping vs production
The security posture that works for prototyping doesn't work for production:
| Aspect | Prototyping | Production |
|---|---|---|
| Credentials | Hardcoded values acceptable | Secret management, rotation |
| Access | Broad permissions for exploration | Least privilege, role-based |
| Change management | Direct deployment | Approval workflows, testing |
| Monitoring | Minimal logging | Comprehensive audit trails |
| Agent access | Full system for experimentation | Restricted to specific tasks |
The challenge is letting people experiment with agents while keeping production systems and sensitive data on approved paths.
Related
- Works with zero data retention (ZDR): with content logging off, Mintive (formerly MintMCP) stores no prompts, responses, or tool-call content.
- Agent Monitor overview
- Agent Monitor rules