Developer Builds Browser Sandbox Prototype For LLM Agents
A developer built Co-do, a browser-based demo testing whether web sandboxing can safely run agentic LLM tasks such as Anthropic's Claude Cowork. The post examines filesystem access (File System Access API), network controls via Content Security Policy, and iframe sandboxing, and notes trade-offs around read/write permissions and exfiltration risks. The analysis assesses feasibility and practical constraints for browser-hosted automation.
Key Points
- 1Demonstrates a Co-do browser demo that runs LLM-driven tasks inside web sandbox constraints
- 2Highlights file, network, and execution isolation as essential controls to prevent data exfiltration
- 3Implies developers can prototype safe agent workflows but must enforce CSP, FS APIs, and iframe sandboxing
Scoring Rationale
Practical, implementation-focused analysis with useful guidance, limited by single-author experimentation and absence of broader validation.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
