NVIDIA launches an open platform to control AI agents
The platform lets users monitor, limit, and stop AI agents that exceed their defined permissions.
NVIDIA announced the Open Agent Safety Platform, an open-source platform for tracking what AI agents do. It lets users set limits and stop agents that try to act beyond the permissions they received. One component, OpenShell, isolates each agent and applies rules to files, networks, tools, and other system resources.
The idea is to have a separate layer oversee the agent. Sentry runs on BlueField-4 hardware and can monitor actions, stop an agent, or place it in quarantine if it exceeds its limits. The platform can also audit actions and decisions and verify who authorized each task. This helps control access, but it does not guarantee that the AI’s decisions are correct.
To get started, look at which permissions each agent receives and which rules are checked before and during its operation. The platform also supports Slack integration to approve or reject requests for additional permissions. Even with these controls, it is important to review the agent’s decisions: limiting its actions does not mean it will always make the right choices.
The link opens the original content; some require a subscription.
Go further on INEMA.CLUB
More news
- 01/10JEV-GW: an AI specialized in helping with decisionsThe JEV-GW presentation includes links to the event page and the project guide.
- 01/10Long-running tasks: a protocol for AI agents to work for hours or daysThe discussion presents a `/goal` prompt protocol for continuous AI agent tasks amid the shift toward autonomous agents.
- 01/10Dot: the ChatGPT assistant that stays active in the cloudThe analysis gathers confirmed facts and issues raised in review videos about OpenAI’s agent for Pro and Business subscribers.
- 28/09agente-claude-codex 1.1.1 updates task records and checksThe update preserves history, checks only what ran, verifies sharing, and keeps memory up to date.
- 23/097 principles to manage AI agentsComparison of external material with INEMA’s methodology for managing AI agents.