Skip to content

Security: vstorm-co/pydantic-deepagents

Security

SECURITY.md

Security Policy

Supported Versions

Version Supported
0.3.x (latest)
< 0.3.0

Only the latest minor release receives security fixes. We recommend always using the latest version.

Reporting a Vulnerability

Please do not report security vulnerabilities via public GitHub Issues.

To report a vulnerability, email the maintainers at:

security@vstorm.co

Include in your report:

  • Description of the vulnerability
  • Steps to reproduce
  • Affected versions
  • Potential impact
  • Any suggested fix (optional)

Response Timeline

Stage Target
Acknowledgement Within 48 hours
Initial assessment Within 5 business days
Fix or mitigation Within 30 days for critical/high
Public disclosure After fix is released

We follow coordinated disclosure — we ask that you give us time to release a fix before public disclosure.

Scope

In scope:

  • Remote code execution
  • Prompt injection via tool inputs or agent context
  • Unsafe deserialization
  • Privilege escalation in multi-agent coordination
  • Docker sandbox escape
  • Sensitive data leakage (MEMORY.md, context files, API keys)

Out of scope:

  • Vulnerabilities in third-party dependencies (report to the respective project)
  • Issues requiring physical access to the machine
  • Social engineering

Security Features

pydantic-deep includes several security-relevant features:

  • pydantic-ai-shields integration — 10 guardrails: prompt injection defense, PII detection, secret redaction, cost controls, tool-level access control
  • Docker sandbox — Sandboxed execution backend for untrusted workloads
  • Budget enforcement — Hard cost limits prevent runaway agent spend
  • Type-safe inputs — All tool inputs validated via Pydantic models
  • 100% test coverage + Pyright/MyPy strict mode — Reduces surface for type-related vulnerabilities

Acknowledgements

We thank all security researchers who responsibly disclose vulnerabilities to us. Confirmed reporters will be credited in the release notes unless they prefer to remain anonymous.

There aren’t any published security advisories