Security researchers at Zenity Labs have discovered a vulnerability in Amazon Web Services allowing a single malicious prompt to compromise multiple artificial intelligence agents. According to a report by The Decoder, a single publicly accessible AI agent operating on Amazon’s Bedrock AgentCore was sufficient to hijack every other AgentCore agent housed within the same AWS account and region. The exploit leveraged an unrestricted internal AWS interface designed to handle temporary cloud credentials, which the agents could access without limitation. Amazon has since addressed the security flaw and applied stricter default permissions to the affected agents.
The Risks of Interconnected AI Infrastructure
As enterprises increasingly deploy autonomous software agents capable of executing multi-step workflows, security architectures face new attack vectors. Multi-agent systems often share underlying cloud environments, IAM roles, and credential management interfaces. When a single entry point lacks proper isolation or least-privilege boundaries, an attacker can pivot laterally across internal network layers. This incident highlights the challenge of securing cloud-native AI frameworks where automated agents require programmatic access to underlying infrastructure APIs.
Unresolved Questions and Future Monitoring
Several details regarding the incident remain undisclosed. The public disclosures and patches did not specify how long the vulnerability existed prior to discovery, whether malicious actors exploited it in the wild before the fix, or the exact scope of affected client accounts. Observers and cloud administrators should monitor future AWS security advisories and review their default agent IAM permissions to ensure proper isolation of autonomous workloads.
AWS has implemented fixes to secure the affected infrastructure, but maintaining rigorous permission boundaries remains critical for multi-agent deployments.
Lead image: panumas nikhomkhai / Pexels
