AWS, Google, and Vercel Agent Flaws Let Attackers Trigger Tools Without Running the Model

Security vulnerabilities in agent infrastructures from AWS, Google, and Vercel allowed attackers to trigger agent tools without model authorization, bypassing system prompts and guardrails. AWS has fixed its AgentCore issue, Google updated ADK to version 2.5.0, and Vercel patched its AI SDK harness packages. These flaws vary in attack conditions and require different attacker capabilities, but are limited to the tools each agent can access.