ML-LabLoop is currently alpha software.

At this stage, security fixes are generally applied to the current development version only. Formal long-term support policies may be introduced after the project reaches stable releases.

Please do not report security vulnerabilities through public GitHub issues.

Report suspected vulnerabilities by email:

alex@cloudcell.nz

Suggested subject:

ML-LabLoop Security Report

Please include as much detail as reasonably possible:

• affected version, commit, or branch
• host operating system and libvirt/QEMU version
• guest image or template version where relevant
• steps to reproduce
• proof-of-concept input, file, script, or command if available
• expected behavior
• observed behavior
• security impact
• whether the issue is public or privately discovered

ML-LabLoop is an early-stage project, so response times are best effort.

The intended process is:

1. Acknowledge the report.
2. Reproduce and assess the issue.
3. Prepare a fix or mitigation where appropriate.
4. Credit the reporter if desired and appropriate.
5. Publish a security note if the issue affects public users.

Security-relevant issues may include, but are not limited to:

• escape from the hostile experiment container into the VM driver zone
• escape from the lab VM to the host
• bypass of the nftables egress fencing applied to the hostile user
• privilege escalation via labloop-exec, labloop-build, or the sudoers rules
• access to trusted MCP state or services from the hostile zone
• exposure of secrets, credentials, or host files inside template images
• persistence of hostile artifacts across template rebuilds or clone resets
• unsafe parsing or execution of experiment artifacts on the driver side

The following are usually out of scope unless they demonstrate a concrete security impact:

• missing hardening in alpha-only development scripts
• issues requiring an already-compromised host or already-root guest
• denial-of-service from intentionally huge inputs without a specific flaw
• reports generated only by automated scanners without analysis
• social engineering or phishing
• risks that are documented and accepted in deploy/SECURITY-MANUAL.md or deploy/SECURITY-COMPLIANCE-NOTE.md, unless a concrete bypass is shown

Please give the maintainers a reasonable opportunity to investigate and fix reported vulnerabilities before public disclosure.

ML-LabLoop is provided without warranty. See the project license for full terms.