Google CodeMender adds sandbox and human-check controls for AI vulnerability agents
The preview update isolates exploit verification in an OS sandbox or VM, requires human approval for local writes and commands, and lets teams scan a repo, folder or only changed files; Auth Manager is generally available.
Why it matters
Google is packaging the same lesson OpenAI's Linux exploit pipeline made vivid: autonomous cyber agents need hard containment before enterprises will run them. CodeMender's sandbox or VM isolation plus human gates on local writes and commands are the enterprise control plane missing from earlier agent demos. Flexible scan scopes also matter for cost when models burn money on full-repo passes. Auth Manager reaching general availability extends the same identity story from coding agents to broader agent credentials.
Key facts
- CodeMender still in preview
- Sandbox or isolated VM for exploit checks
- Human approval for local writes and commands
- Scan repo, folder, or changed files only
- Auth Manager generally available