Researchers recently discovered security weaknesses in artificial intelligence (AI) systems used by Amazon Web Services (AWS), Google, and Vercel. These flaws could allow attackers to make an AI agent perform actions without first getting approval from the AI model itself. Normally, an AI reviews a request, decides whether an action is safe, and then carries it out. In these cases, attackers found ways to skip that decision-making process, meaning the AI’s built-in safety checks were never used.

The affected products included Amazon Bedrock AgentCore, Google’s Agent Development Kit (ADK), and Vercel’s AI SDK. Although the vulnerabilities were different for each company, they all shared a similar problem: the software trusted certain instructions without confirming they actually came from the AI. This could allow someone with the right level of access to trick an AI agent into using tools it was allowed to access.

The good news is that the vulnerabilities have already been fixed. AWS updated its managed service automatically, Google released a security update in ADK version 2.5.0, and Vercel patched the affected versions of its AI software. Companies using these tools should install the latest updates to stay protected.

Researchers emphasized that these flaws did not automatically give attackers unlimited access. Any damage would depend on what the AI agent was already allowed to do. For example, if an AI only had permission to answer questions, the risk would be limited. However, if it could access sensitive information, deploy software, or make