Major Cloud Providers Patch Critical AI Agent Vulnerability Bypassing Safety Checks

SAN FRANCISCO — Amazon Web Services, Google and Vercel disclosed a critical security vulnerability on Wednesday affecting artificial intelligence agent infrastructure that allowed attackers to forge tool calls without model authorization. The flaw enabled malicious actors to bypass standard safety checks by injecting untrusted instructions directly into execution tools.

The issue stems from insufficient input validation within specific agent workflows. In affected configurations, the logic permitted certain flows to skip the central AI model entirely. This architectural gap allowed forged commands to reach underlying systems without undergoing the scrutiny typically applied by large language models designed to filter harmful content or unauthorized actions. Security researchers noted that an attacker could theoretically manipulate these agents into executing arbitrary code, accessing sensitive data, or performing financial transactions under false pretenses.

The vulnerability was identified in infrastructure supporting AI agents across all three major platforms. AWS confirmed the issue affected specific implementations of its Bedrock agent framework. Google stated the flaw impacted certain configurations within Vertex AI Agent Builder. Vercel reported that the defect existed in serverless functions handling automated agent logic on their platform. All three companies have since issued patches to address the input validation gaps and enforce mandatory model authorization for all tool invocations.

The discovery highlights emerging risks as organizations increasingly deploy autonomous agents capable of interacting with external APIs, databases, and cloud resources without human intervention. Unlike traditional software vulnerabilities that rely on memory corruption or injection attacks against a single application, this flaw exploited the trust boundaries between an AI's reasoning layer and its execution environment. By circumventing the model itself, attackers could bypass safety guardrails trained to prevent harmful outputs.

Industry experts warn that similar logic errors may exist in other agent frameworks not yet audited for this specific failure mode. The incident underscores the complexity of securing autonomous systems where the distinction between a user prompt and an internal system command can be blurred by sophisticated input manipulation.

While patches are now available, cloud providers have advised customers to review their agent configurations immediately to ensure updated versions are deployed. Questions remain regarding whether any active exploitation occurred prior to the public disclosure on Wednesday. Neither AWS nor Google has confirmed if they detected unauthorized tool executions in production environments before mitigating the flaw. Vercel stated it is investigating potential impact but declined to comment on specific incidents.

As AI agents become more integrated into enterprise workflows, ensuring that every action passes through authorized safety layers remains a critical challenge for developers and security teams alike.