A newly disclosed attack chain in Amazon Bedrock AgentCore lets attackers with chat access to an exposed AI agent steal temporary cloud credentials and compromise other agents in the same AWS account and region.
The research, named AgentCorruption, connected prompt injection, metadata access, and excessive execution-role permissions.
AgentCore runs containerized agents inside Firecracker microVMs. Researchers showed that an agent with HTTP or shell tools could be tricked by a malicious prompt into contacting the metadata endpoint at 169.254.169.254.
AWS Bedrock AgentCore Flaw
The resulting server-side request forgery exposed credentials associated with the agent’s execution role. The returned credentials included an access key ID, secret access key, and session token.

Researchers exported them to a separate machine and verified that AWS recognized subsequent requests as originating from the compromised agent’s assumed role. Further exploitation therefore used legitimate AWS APIs without requiring continued interaction with the agent.
The execution role Zenity examined was not restricted to the compromised agent. Its permissions extended across AgentCore resources within the account and region, turning credential theft into a broader compromise.
Using logs:DescribeLogGroups, researchers discovered agent identifiers from CloudWatch log names. Broad Elastic Container Registry permissions then let them retrieve other agents’ container images and inspect application code.
The bedrock-agentcore:InvokeAgentRuntime permission also enabled access to internal agents that the attacker could not otherwise reach. Memory permissions compounded the exposure.
ListActors and ListSessions revealed users and sessions, while ListEvents exposed stored conversation content. CreateEvent and DeleteEvent enabled attackers to manipulate conversation history, potentially changing an agent’s decisions or removing contextual evidence.
The supplied overview additionally describes persistent memory poisoning against agents with long-term memory enabled. Attackers could inject events that influenced future sessions.

Permissions including bedrock-agentcore:GetResourceApiKey and secretsmanager:GetSecretValue also exposed credentials intended for connected tools and external services. Zenity’s disclosure timeline states that the initial metadata report was submitted on December 25, 2025.
AWS subsequently reported that newly deployed agents used IMDSv2 beginning February 14, 2026. Researchers observed substantial restrictions to the default execution role during their September 29 review.
AWS characterized execution-role credential access through metadata as documented and expected behavior, rather than an unexpected capability. Its documentation warns that any code or actor inside the microVM can retrieve these credentials, independently of network mode.
The distinction matters: metadata access alone does not establish unrestricted cloud compromise. The demonstrated blast radius depended on the permissions attached to the stolen identity.
AWS recommends least-privilege execution roles and avoiding roles more powerful than the users permitted to invoke the associated agent.
Join 16,000+ SOC teams using ANY.RUN to streamline threat investigations and reduce manual effort. Explore for your team
The post AWS Bedrock AgentCore Flaw Let Attackers Hijack AI Agents and Steal Cloud Credentials appeared first on Cyber Security News.
Discover more from RSS Feeds Cloud
Subscribe to get the latest posts sent to your email.

You must be logged in to post a comment.