Sandboxes won't save you from OpenClaw
5 hours ago
- #Agent Permissions
- #AI Security
- #Sandbox Limitations
- AI agent misbehavior is a permissions problem, not a sandbox issue.
- OpenClaw has caused significant issues like deleting inboxes and spending large sums in crypto.
- Sandboxes isolate workloads but don't prevent agent misbehavior with third-party services.
- Agents need granular permissions, not just sandboxing, to be safe.
- Current permissions like OAuth are too coarse for agent needs.
- Proposed solutions include pre-approving contacts for emails and limited credit card access.
- New interfaces and standards are needed for agent security.
- Finance is likely the first sector to adopt these new standards.