[Evaluation] Monty (Pydantic) — Python Code Sandbox for Layer 5 Tool Restrictions

[Steffen025]

Summary

What: Evaluate Monty (Pydantic's sandboxed Python interpreter) for enabling safe Python code execution in pai-collab's untrusted contribution zone.

Why: Layer 5 (Tool Restrictions) currently blocks Python execution entirely. Monty would allow agents to validate, test, and review Python contributions from external contributors without security risk.

Impact: Unlocks Python as a first-class contribution type while maintaining 6-layer defense model.

Technology

Monty — A minimal, secure Python interpreter written in Rust:

• 0.06ms startup (vs Docker ~195ms)
• Strict sandbox: no filesystem, network, or environment access
• Resource limits: memory, time, stack depth
• External functions: host-controlled only
• Serialization: pause/resume execution state

Source: https://github.com/pydantic/monty
Docs: https://docs.pydantic.dev/monty/

Full Evaluation

📄 Complete evaluation: research/2026-02-10-monty-technology-evaluation.md

Use Cases for pai-collab

1. Python Contribution Validation — Agents can execute and validate Python code from untrusted contributors
2. Review Mode Enhancement — Extend review-mode with safe Python execution
3. CI/CD Python Gate — Automated validation for .py files in PRs

Proposed Next Steps

Phase 1: Evaluation (This Issue)

• Create proof-of-concept in contributions/review-mode
• Test with synthetic Python contributions
• Measure startup time, memory, detection accuracy

Phase 2: Integration (Future PR)

• Add Monty dependency to review-mode
• Implement validatePython() utility
• Update review-format.md SOP
• Add CI gate for Python files

Phase 3: Documentation

• Update TRUST-MODEL.md — Monty as Layer 5 enabler
• Create SOP: python-contribution-validation.md

Questions for Maintainers

1. Is this aligned with pai-collab's security philosophy?
2. Should we prioritize Phase 1 evaluation?
3. Who should be assigned for PoC implementation?

Labels

type/research, security, layer-5, python, evaluation

Effort Estimate

4-6 hours for Phase 1 PoC

References

• Related projects: pai-content-filter, pai-secret-scanning
• Trust model: TRUST-MODEL.md
• Requested by: @Steffen025 (via Jeremy agent)