Back to browse
GitHub Repository

Hermetic blinded agents

2 starsPython

Hermetic 'blinded' agents using Starlark sandbox

by jakewins·Mar 1, 2026·2 points·0 comments

AI Analysis

●●●BangerBig BrainWizardryZero to One

Starlark sandbox + opaque values defends agents from prompt injection, not just wrapping smolagents.

Strengths
  • Novel security model: opaque values force agents to manipulate data without reading it
  • In-process Starlark sandbox avoids container/VM overhead while maintaining hermetic execution
  • Directly addresses real agent failure mode (prompt injection via untrusted PDFs, emails, etc.)
Weaknesses
  • Unaudited PoC status means risky for production use today
  • Limited examples of blinded agent patterns; unclear how complex real workflows map to the API
Target Audience

AI/ML engineers building agentic systems with untrusted data inputs

Similar To

smolagents · LangChain agents · Anthropic's Secure Compute

Similar Projects

AI/ML●●Solid

A self upgrading agent that learns from failure

Agent writes its own Python tools and saves rules to avoid repeating mistakes.

WizardryBold Bet
grimm76321
1029d ago