agent-security
4 articles
How Anthropic Contains Claude: Agent Safety Is Not Just Asking for More Confirmations
Anthropic explains how claude.ai, Claude Code, and Claude Cowork contain agents: model defenses miss, permission prompts create fatigue, and the hard boundary is the VM, sandbox, filesystem policy, and egress control.
Every Agent Needs a Bouncer: Brex Open-Sources CrabTrap, an LLM-Judge HTTP Proxy for Production Agents
Brex open-sourced CrabTrap, an HTTP proxy for agent requests. Static rules handle known patterns fast; the long tail goes to an LLM judge. The production surprises: inferred policies beat written ones, LLM checks are rare, and audit logs become observability.
Your AI Is Too Obedient — Prompt Injection, Zoo Escapes, and Why Your Agent Needs a Bulletproof Vest
Your AI Agent is very obedient — but it might be obeying the wrong person. Prompt Injection is social engineering for AI. Tool Use Exploitation is giving a Swiss Army knife to a 5-year-old. Context Poisoning is someone secretly changing books in a library. And then there's the zoo escape.
How Dangerous Is the MCP You Use Every Day? A Paper Dissects 12 Security Landmines in AI Agent Protocols
New paper: comprehensive security threat modeling of MCP, A2A, Agora, ANP (4 major AI agent protocols). Finds 12 protocol-level risks, including MCP being tricked 73.3% into calling wrong tool providers. Important for Claude Code, OpenClaw, Cursor users.