Launching today
NEXO Privilege Gate
A least-privilege security gateway for AI agents
1 follower
A least-privilege security gateway for AI agents
1 follower
NEXO Privilege Gate v0.4.0 is an open-source least-privilege gateway for AI agents. It now enforces structured filesystem actions inside a local sandbox, blocks path escapes and sensitive-file aliases, requires human approval for destructive actions, quarantines deletions, and records a tamper-evident audit trail. GPT-6 Astra also found and helped fix a real path-alias authorization bypass before release.
I built NEXO Privilege Gate around a simple question: if an AI agent can request an action on your computer, should that automatically mean it has permission to execute it?
v0.4.0 moves NEXO beyond a browser policy demo into real local filesystem enforcement inside an isolated sandbox.
The current build supports ALLOW, REVIEW, and DENY decisions, explicit human approval for destructive actions, quarantine instead of permanent deletion, path-traversal protection, sensitive-file protection, and a tamper-evident audit chain.
During the final pre-release review, GPT-6 Astra found a real authorization bypass: alternate path representations such as `.env/.` and Windows-specific `.env.` could cause policy evaluation and filesystem resolution to disagree about the same protected target.
Astra helped implement a canonical-path fix and added eight HTTP regression tests. The complete suite now passes 34/34 tests.
The scope is still intentionally narrow: NEXO currently protects filesystem operations inside its sandbox, not arbitrary Windows, browser, network, financial, or MCP actions.
The long-term goal is a reusable authorization layer between increasingly capable AI agents and the systems they operate.
Feedback on the threat model, policy design, and the next tool domains to protect is very welcome.