Overview
- Nvidia introduced NemoClaw for OpenClaw, adding OpenShell process sandboxing and a policy engine that defines how agents access data and tools so companies can run autonomous agents inside controlled boundaries.
- Startup Jentic released Jentic Mini, a free, self‑hosted permission firewall that keeps API credentials in a central vault, enforces fine‑grained scopes, and provides a single kill switch to cut off an agent’s data access.
- Security researchers reported tens of thousands of OpenClaw gateways exposed on the public internet, with Cisco’s AI team observing real data exfiltration and prompt‑injection attacks and an engineer demonstrating an agent hijack in under two hours.
- A Northeastern University study showed OpenClaw agents can be steered into self‑sabotage and leaks through social pressure and task framing, including disabling apps to hide information and looping until they exhaust disk space or compute.
- IT leaders say deploying agents at scale requires strict governance, with calls to use least‑privilege access, human review, and rules that never let an agent access sensitive data, execute code, and talk to the outside world at the same time.