English
Español
Français
Deutsch
日本語
한국어
العربيةPass role to Bedrock AgentCore
Can create a Bedrock AgentCore interpreter with a role and run Python as it.
Grants escalation only in combination. All of these are needed together:
iam:PassRolebedrock-agentcore:CreateCodeInterpreterbedrock-agentcore:StartCodeInterpreterSessionbedrock-agentcore:InvokeCodeInterpreteriam:PassRole can create a Bedrock AgentCore interpreter with a role and run Python as it, but only when it is held together with bedrock-agentcore:CreateCodeInterpreter, bedrock-agentcore:StartCodeInterpreterSession, bedrock-agentcore:InvokeCodeInterpreter. Held together they reach administrator access.
Why it is dangerous
A principal holds iam:PassRole on the target role plus bedrock-agentcore:CreateCodeInterpreter, StartCodeInterpreterSession and InvokeCodeInterpreter, so it can create a Bedrock AgentCore code interpreter with the target role attached, open a session and invoke attacker-supplied Python, that Python runs in the AgentCore sandbox with the target role's permissions and reads the target role's live credentials from the MicroVM metadata service. Because the target role trusts bedrock-agentcore.amazonaws.com the pass is accepted; if the target role is privileged this is full privilege escalation via a new 2025 AI-runtime sink, which is why it is marked needs-verification.
In MITRE ATT&CK terms this maps to T1648, Serverless Execution, under the privilege escalation tactic. AWS records it in CloudTrail as CreateCodeInterpreter, StartCodeInterpreterSession, InvokeCodeInterpreter. The call shows up in CloudTrail but blends into normal administration, so it needs a rule to stand out.
Detection
CreateCodeInterpreterStartCodeInterpreterSessionInvokeCodeInterpreterPrerequisites
Target role's trust policy must allow bedrock-agentcore.amazonaws.com. All three bedrock-agentcore actions (create, start-session, invoke) plus iam:PassRole are required; the code interpreter runs the attacker's Python as the agent role, so no separate cred-theft step is needed.
MITRE ATT&CK
T1648 · Serverless ExecutionTactic
Privilege Escalation
A control that would remove it: Remove iam:PassRole for this role or scope it with iam:PassedToService.
Whether to apply it depends on how the permission is used in your environment, so it is a business decision rather than an automatic fix. The same permission can be load-bearing in one account and pure risk in another.
Pathbreak confirms whether this path resolves in your account, ranks it against the others by how much it opens up, and uses your threat model and business context to show which paths are worth cutting.
Questions
What does iam:PassRole allow?
Can create a Bedrock AgentCore interpreter with a role and run Python as it.
What else is needed for iam:PassRole to be dangerous?
It becomes an escalation only when held together with bedrock-agentcore:CreateCodeInterpreter, bedrock-agentcore:StartCodeInterpreterSession, bedrock-agentcore:InvokeCodeInterpreter. On its own it is not enough, which is what makes the combination easy to miss.
How is pass role to Bedrock AgentCore detected?
Watch for CreateCodeInterpreter, StartCodeInterpreterSession, InvokeCodeInterpreter in CloudTrail. The call shows up in CloudTrail but blends into normal administration, so it needs a rule to stand out.
How do you break this path?
A control that would remove it is: Remove iam:PassRole for this role or scope it with iam:PassedToService. Whether to apply it depends on how the permission is used in your account, so it is a business decision rather than an automatic fix.
Which MITRE ATT&CK technique is this?
T1648, Serverless Execution, under the privilege escalation tactic.
Related paths
References
Find this in your own account
We run a read-only scan of one AWS account and show you which of these paths exist, ranked by blast radius, with the change that breaks each one. First validated paths within the hour.