ClawAudit verdict
operating-autodl-training
autodl-train
SSH-based remote training management that explicitly constrains operations to a configured project directory and includes documented safety rules refusing destructive commands; all capabilities match the stated purpose of managing ML training jobs.
⚠ Flagged for review — coarse, uncorroborated signal, not a confirmed exploit. Review the config yourself before installing.
Automated static analysis — not a human review. ClawAudit flags capabilities, not confirmed intent, and can produce false positives. Disagree with this verdict? Use Dispute below.
Findings (6)
Possible hardcoded credential
references/usage.md · code · PASSWORD='your-server-password
Pipe to bash — executes piped content as shell commands
scripts/common.py · prose · downgraded · | bash
References SSH/GPG private keys
config.example.json · prose · downgraded · ssh_key
Accesses .ssh directory
config.example.json · prose · downgraded · .ssh/
subprocess execution — runs system commands from Python
scripts/common.py · prose · downgraded · subprocess.run(
Sets world-executable permissions
scripts/remote_train.py · prose · downgraded · chmod 700
Why the tier is capped
Execution sink present in raw bytes (Hard Floor: class A/D/F). Final tier capped at Caution — cannot be lifted by any downgrade, example-payload opt-in, or allowlist.
Permissions & capabilities
No declared permissions — minimal attack surface.
Is this flag fair?
Thanks — recorded.