AgentCorruption: Zenity Labs reports AWS AgentCore flaws
Zenity Labs reported AgentCorruption on 8 October 2026, a chain of Amazon Bedrock AgentCore flaws. A single prompt to one public-facing agent gave access to all AgentCore agents in the same AWS account and region, plus conversations, memories and credentials. Zenity says AWS made IMDSv2 the default and reduced the default role's permissions.
Facts read on .
Zenity Labs presented AgentCorruption at SecTor 2026 in Toronto and published a press release on 8 October 2026. The research covers a chain of flaws in Amazon Bedrock AgentCore. Using one prompt to a public-facing agent that had an outbound-request tool, researchers reached credentials for a default IAM role whose permissions covered every AgentCore agent in the same AWS account and region.
From there Zenity says it could invoke agents it was not authorized to use, read private conversations and long-term memories, download container images and source code, and retrieve API keys and OAuth tokens held in AWS Secrets Manager. It also planted memories that sent future conversations to an outside destination, a persistence route that outlives the first session.
Zenity says it reported the findings to AWS on 25 December 2025. Following that report, AWS made IMDSv2 the default for AgentCore deployments, and Zenity's testing confirmed AWS had cut the default execution role's permissions. The identity lesson is plain: one over-broad role shared by many agents turns a single compromised agent into a regional problem. The account is Zenity's own.
Sources
- Zenity Labs reports AgentCorruption, a chain of AWS AgentCore flaws (zenity.io), read