How we contain Claude across products

kept by eddie may 27

Anthropic details how it contains AI agent blast radius across three products using sandboxes, VMs, and human-in-the-loop controls, while documenting missed risks like pre-trust code execution and user-mediated prompt injection.

GitHub - anthropics/defending-code-reference-harness: Skills for threat modeling, scanning, triage, patching, plus an autonomous scanning harness you can /customize
Anthropic released an open-source reference harness and Claude Code skills for automating vulnerability discovery, triage, and patching using Claude, alongside a managed product called Claude Security.
also on gVisor, Claude, Claude Code, Anthropic, #developer-tools
A field guide to sandboxes for AI
Sandboxing tools for AI agents differ fundamentally by boundary (containers, gVisor, microVMs, Wasm), not just by policy, and picking the wrong one creates either leaks or excessive cost.
also on gVisor, #security, #ai-agents, #developer-tools
Claude Cowork exfiltrates files
Researchers demonstrate file exfiltration from Claude Cowork via prompt injection, exploiting an acknowledged but unremediated vulnerability.
also on Claude Cowork, #ai-safety, #security, Anthropic, #ai-agents
First impressions of Claude Cowork, Anthropic’s general agent
Claude Cowork wraps Claude Code in a sandboxed UI for non-developers, offering general agent capabilities but raising prompt injection concerns.
also on Claude Cowork, #security, Claude Code, Anthropic, #ai-agents, #developer-tools
Claude Code and What Comes Next - by Ethan Mollick
Claude Code, using agentic harness and compaction, autonomously built a working website and game, showing a major leap in AI capability for sustained work.
also on Model Context Protocol, Claude Code, Anthropic, #ai-agents, #software-engineering, #developer-tools