Anthropic's Zero Trust for AI Agents Sets the Right Test. The Bearer Token Fails It¶

📊 Level ⭐ | 3.2KB

相关实体¶

• microsoft is quietly shopping for an openai replacement
• vietnam to develop domestic cloud
• akamai acquires israeli ai browser security startup layerx f

→ 原文存档

核心要点¶

1. Anthropic's policy paper on AI agent security — Anthropic published a position paper advocating Zero Trust architecture for AI agents: every tool call should be authenticated, authorized, and audited independently, not relying on session-based trust.
2. The bearer token problem — The author argues that long-lived bearer tokens (OAuth access tokens, API keys) are fundamentally incompatible with agentic AI. Agents accumulate tokens across services; if one token is compromised, the attacker has the agent's full delegated authority. The paper calls for short-lived, scoped, refresh-able credentials.
3. The test the bearer token fails — A Zero Trust agent should NOT be able to perform an action it wasn't specifically authorized for in the current scope, even with a valid token. Current bearer token models fail this test: a valid repo:read token can sometimes be escalated to repo:write via confused deputy attacks.
4. Practical recommendation: per-action credential injection — Instead of giving an agent a token at session start, the runtime should fetch a fresh, narrowly-scoped credential for each tool call, ideally with user-in-the-loop confirmation for high-risk actions (delete, send-money, modify-permissions).

实战启示¶

• 生产级实施建议：基于上述要点，构建可落地的实施方案
• 风险评估：在采用前评估安全/性能/成本 trade-off
• 参考架构：借鉴同领域最佳实践，避免常见陷阱

上线状态 / 链接¶

• 原文链接: https://blog.hello.coop/2026/06/anthropics-zero-trust-for-ai-agents-sets-the-right-test-the-bearer-token-fails-it/
• 作者/平台: newsletter
• 类型: newsletter / 行业分析