Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-112408 Overview
CVE-2026-112408 is a critical vulnerability in the Amazon SageMaker Python SDK affecting the default_bucket() function. The SDK fails to verify ownership of the S3 bucket it uses by default, enabling an attacker to pre-create a predictably named bucket in a target account's expected namespace. When a victim later invokes default_bucket(), the SDK trusts the attacker-controlled bucket and reads or writes training artifacts from it. This leads to code execution in the SageMaker execution environment and exposure of the associated IAM credentials. The flaw is categorized as [CWE-940] Improper Verification of Source of a Communication Channel.
Critical Impact
Remote actors can achieve code execution inside the target SageMaker execution environment and obtain IAM credentials by hijacking a predictably named S3 bucket used without ownership validation.
Affected Products
- Amazon SageMaker Python SDK V3 before 3.10.0
- Amazon SageMaker Python SDK V2 before 2.257.3
- sagemaker package distributed via PyPI
Discovery Timeline
- 2026-10-21 - CVE-2026-112408 published to NVD
- 2026-10-22 - Last updated in NVD database
Technical Details for CVE-2026-112408
Vulnerability Analysis
The Amazon SageMaker Python SDK exposes a helper named default_bucket() that returns the name of an S3 bucket used to stage training data, model artifacts, and intermediate outputs. The SDK generates this bucket name deterministically based on account and region context. Before versions 3.10.0 and 2.257.3, the SDK calls the bucket by name without confirming that the bucket in the current account is actually owned by the caller. An external actor who registers the predictably named bucket in their own AWS account before the victim does becomes the owner of the staging channel used by SageMaker jobs.
SageMaker training and processing jobs load code and configuration from this staging bucket. Once the attacker controls that bucket, they can place malicious scripts or pickled artifacts there. When the SDK pulls these artifacts into the execution environment, the attacker achieves arbitrary code execution under the IAM role attached to the job and can exfiltrate temporary credentials for that role.
Root Cause
The root cause is missing verification of the source of a communication channel. The SDK trusts bucket name resolution without issuing an s3:GetBucketAcl or s3:HeadBucket check against the expected account ID. Predictable naming combined with absent ownership validation enables a cross-account bucket squatting attack.
Attack Vector
An unauthenticated remote attacker pre-creates the predictable bucket name in an AWS account they control. When a victim running a vulnerable SDK version invokes default_bucket(), the SDK uses the attacker's bucket for data staging. The attacker then plants payloads retrieved by SageMaker jobs, gaining code execution and IAM session credentials. No user interaction on the victim side beyond normal SDK usage is required.
See the GitHub Security Advisory GHSA-5rx8-8fr2-f855 for the vendor's technical description.
Stop Remote Code Execution Before It Runs
Singularity™ Endpoint detects and blocks exploit attempts at machine speed, including zero-days and novel attack techniques.
Stop Exploit AttemptsDetection Methods for CVE-2026-112408
Indicators of Compromise
- SageMaker jobs reading from or writing to S3 buckets owned by an AWS account ID that does not match the job's account.
- Unexpected AssumeRole or GetSessionToken activity originating from SageMaker execution roles followed by API calls from unfamiliar source IPs.
- Creation of buckets matching the pattern used by default_bucket() in third-party accounts before the victim account provisions them.
Detection Strategies
- Inventory installed sagemaker package versions across data science workstations and build pipelines and flag any version below 3.10.0 or 2.257.3.
- Enable AWS CloudTrail data events for S3 and correlate PutObject and GetObject operations issued by SageMaker roles against expected bucket ownership.
- Alert on SageMaker job logs that reference buckets outside the organization's approved bucket list.
Monitoring Recommendations
- Monitor IAM Access Analyzer findings for SageMaker execution roles accessing external accounts.
- Track GuardDuty findings in the CredentialAccess and Exfiltration categories tied to SageMaker roles.
- Audit S3 bucket creation events for names that follow the sagemaker-{region}-{account-id} convention in unexpected accounts.
How to Mitigate CVE-2026-112408
Immediate Actions Required
- Upgrade the sagemaker package to 3.10.0 (V3) or 2.257.3 (V2) in all notebooks, pipelines, and container images.
- Pre-create the default SageMaker bucket in every region and account you operate in to prevent squatting of the predictable name.
- Rotate IAM credentials for any SageMaker execution role suspected of interacting with an attacker-controlled bucket.
Patch Information
Amazon released fixed versions available on PyPI as sagemaker 3.10.0 and sagemaker 2.257.3. Corresponding source releases are published at GitHub Release v3.10.0 and GitHub Release v2.257.3. Additional guidance is published in AWS Security Bulletin 2026-142.
Workarounds
- Explicitly pass a controlled, pre-provisioned bucket name to SageMaker SDK calls instead of relying on default_bucket().
- Apply least-privilege S3 bucket policies that restrict SageMaker roles to buckets owned by the organization using aws:ResourceAccount and s3:ResourceAccount conditions.
- Deploy Service Control Policies (SCPs) that deny SageMaker roles from accessing S3 buckets outside approved account IDs.
# Upgrade the SageMaker Python SDK to a fixed version
pip install --upgrade "sagemaker>=3.10.0"
# Or, for the V2 release line:
pip install --upgrade "sagemaker>=2.257.3,<3.0.0"
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

