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

Security flaws have been found in the agent infrastructure of Amazon Web Services, Google, and Vercel, allowing unauthorized instructions to access an agent's tools without verification. These flaws enable attackers to trigger tools without running the associated model, bypassing system prompts and content filters. In some cases, the model is not run at all, preventing model-level guardrails from intervening. The vulnerabilities affect products from Amazon, as well as those from Google and Vercel. This allows untrusted or forged instructions to reach the tools without proper authorization. The flaws in these products can be exploited to execute unauthorized actions.