Dev.to Security πŸ” Cybersecurity πŸ‘ 0

Three agent-sandbox releases this week, and the question they leave open

Disclosure: I help the founder of RayTrace, which is mentioned at the end. What shipped External sandboxes (Simon Willison / Prime Radiant): the agent stays on your machine, commands run in a Linux guest. N

Disclosure: I help the founder of RayTrace, which is mentioned at the end.

What shipped

  • External sandboxes (Simon Willison / Prime Radiant): the agent stays on your machine, commands run in a Linux guest. Note: Claude Code sends only Bash there; Read/Edit stay on the host.
  • Eclipse Enclave 1.0: MIT sandbox for Claude Code, Codex, OpenCode and more, with a network allowlist and enclave network log.
  • Claude Code Code Mode: Claude writes JavaScript that makes many tool calls in one step. In auto mode the classifier reviews the code once, not each call.

The common theme

Limit reach without relying on the model's judgment.

The open question

What did the agent actually do inside those limits, and does it match what it told you?

One way to answer it

(Describe RayTrace: reported vs observed, mismatch flags, fork, search. Add a screenshot.)

https://github.com/Godcreatebugs/raytrace-cli

πŸ“° Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.