🟠 High | Source: The Hacker News
Security researchers discovered flaws in AI agent infrastructure from AWS, Google, and Vercel that allowed attackers to trigger tool calls without a language model ever processing the request. Because the model never ran, safety controls such as system prompts, content filters, and guardrails were entirely bypassed. This means an attacker could invoke agent tools — potentially with real-world side effects — purely through forged or untrusted instructions.
Security Architect’s Take: Audit any AI agent pipelines built on AWS Bedrock Agents, Google Agent Builder, or Vercel AI SDK to ensure tool invocations are gated on verified model-turn authorisation, not just incoming request structure. Apply patches immediately and implement independent authorisation checks at the tool-execution layer that do not rely solely on model-level guardrails.
Original advisory: AWS, Google, and Vercel Agent Flaws Let Attackers Trigger Tools Without Running the Model