rwxmap
A tool that provides rwx (read/write/execute) permission classification labels for AI Agent operations. By labeling the reversibility of operations, it helps assess the risk of Agent actions. It supports two modes: pre-annotation of services and real-time evaluation of calls.
Timeline (last 90 days)
A developer on Reddit proposed a project called rwxmap, used to tag AI Agent action permissions with rwx labels to distinguish whether operations are reversible.
rwxmap borrows the Unix rwx permission model, marking r as read-only operations, w as recoverable changes, and x as operations that cannot be undone.
The key difference between rwxmap and the file system permission model is that it does not treat r, w, x as permission bits but as property labels of the operation, shifting from describing the subject's capabilities to describing the object's effects.
rwxmap supports two usage modes: pre-labeling all operations of an entire service, or evaluating each Agent call in real time.
rwxmap can be used together with the developer's other project bareguard: rwxmap labels the call, and bareguard decides whether the Agent may execute that call.
In a blind test on 4,279 operations from unseen services, rwxmap used alone achieved about 82% accuracy, and reached 93% when combined with Jev TypeSafe AI.
rwxmap chooses the stricter label when uncertain (fail-safe principle), and overly permissive errors account for less than 1% of results.
The developer's self-assessment of rwxmap is: "Under-specified, held together by guardrails. A starting point, not a standard."
Both rwxmap and bareguard are open source and available on GitHub (github.com/hamr0/rwxmap and github.com/hamr0/bareguard).
All Facts (9)
Both rwxmap and bareguard are open source and available on GitHub (github.com/hamr0/rwxmap and github.com/hamr0/bareguard).
50%Unverifiedrwxmap borrows the Unix rwx permission model, marking r as read-only operations, w as recoverable changes, and x as operations that cannot be undone.
50%UnverifiedThe key difference between rwxmap and the file system permission model is that it does not treat r, w, x as permission bits but as property labels of the operation, shifting from describing the subject's capabilities to describing the object's effects.
50%Unverifiedrwxmap supports two usage modes: pre-labeling all operations of an entire service, or evaluating each Agent call in real time.
50%Unverifiedrwxmap can be used together with the developer's other project bareguard: rwxmap labels the call, and bareguard decides whether the Agent may execute that call.
50%UnverifiedIn a blind test on 4,279 operations from unseen services, rwxmap used alone achieved about 82% accuracy, and reached 93% when combined with Jev TypeSafe AI.
50%Unverifiedrwxmap chooses the stricter label when uncertain (fail-safe principle), and overly permissive errors account for less than 1% of results.
50%UnverifiedThe developer's self-assessment of rwxmap is: "Under-specified, held together by guardrails. A starting point, not a standard."
50%UnverifiedA developer on Reddit proposed a project called rwxmap, used to tag AI Agent action permissions with rwx labels to distinguish whether operations are reversible.
50%