Skip to content

Security: amitverma-cf/agent.cpp

SECURITY.md

Security Policy

Reporting a Vulnerability

Report a security vulnerability in this project privately. Do not disclose it as a public issue. This gives the maintainers time to prepare a fix before public exposure.

Report it privately with GitHub's private security advisory form: https://github.com/amitverma-cf/agent.cpp/security/advisories/new

Include the version or commit, the platform, the engine and a way to reproduce the problem. Do not open a public issue or pull request for it.

Covered Topics

Vulnerabilities within the following areas are considered high priority:

  • Core Logic: Memory safety issues in the Arena allocator, turn executor or graph runtime.
  • Hardware Abstraction: Vulnerabilities in the abstraction layer that could lead to privilege escalation or unauthorized system access.
  • Tool Execution: Insecure handling of tool parameters that could lead to command injection.
  • Input Sanitization: Failures to properly sanitize or validate outputs before action execution.

Using the Framework Securely

Untrusted Models & Inputs

Autonomous agents are inherently susceptible to Prompt Injection. This project provides the orchestration layer, but it is the developer's responsibility to:

  • Sandbox Tool Execution: Always execute the agent in an isolated environment (e.g., a container or VM) when tools have access to the file system or network.
  • Validate Observations: Sanitize all data returned from tools before feeding them back into the context.
  • Restrict Tool Access: Use the principle of least privilege when registering tools.
  • OS sandbox: on Linux the engine host and run_command run under Landlock + seccomp + rlimits by default (Config::engine_sandbox, Config::tool_sandbox), MCP servers opt in per server. It narrows what a compromised child can touch but is not a substitute for a container or VM; on Windows, MCP servers and run_command run in an AppContainer (tested on Windows 11), and engine hosts get a Job Object only. macOS: a Seatbelt profile is applied but untested; use a VM or container. See docs/security-sandbox.md.

Data Privacy

Since the framework uses a contiguous Arena allocator, sensitive data from previous sessions could remain in memory if the arena is not explicitly zeroed out. While the default behavior is a pointer reset, production deployments should consider zero-filling the arena between sessions if handling high-sensitivity data.

Multi-Tenant Environments

In multi-tenant environments, ensure strict isolation between memory arenas. Do not share arenas between different users or agents.

There aren't any published security advisories