Topic

Code-agent sandboxing

Restricting the libraries and network access available to an agent that executes generated Python, and what happens when a small model tries to escape.

Current clusters

invest1 distinct publisher

Anthropic trades 84% of Claude Code's permission prompts for a working-directory fence

Seatbelt on macOS and bubblewrap on Linux now stop the agent's shell at the folder it was invoked in and route its traffic through a domain allowlist. That is a boundary a security team can approve once.

Publishers:cryptobriefing.com

Reality

Evidence34
Adoption22
Hype gap+20
Incentives66

build1 distinct publisher

213 seconds per agent step evicts hybrid RAG from the local CPU

The FAISS-plus-BM25 retrieval in this writeup does address vocabulary mismatch, but the agent loop around it ran at over 213 seconds a step on CPU, and that figure decided the deployment, not the retrieval design.

Publishers:dev.to

Reality

Evidence34
Adoption14