QuietOS
An outbound permission-control component for personal AI agents that holds the agent's permissions rather than its identity. It uses temporary permits with expiration times and an append-only ledger based on SHA-256 hashes to record approvals and decisions, providing an auditable authorization gate.
Timeline (last 90 days)
A developer shared the QuietOS+Sentinel approach on Reddit, in which QuietOS handles outbound permission control and Sentinel handles inbound content review, with the final decision left to humans.
In QuietOS's design, the Agent holds only identity while QuietOS holds the permissions; approvals expire, and there are verifiable logs.
QuietOS's proof log records every hold and decision with SHA-256 hashes in an append-only ledger, which can be audited afterwards.
The project has released its code publicly, published version 0.1.0 on PyPI, and passes 34/34 of its benchmark test suite.
The developer explicitly stated that there is no monthly fee, and there is not yet a mobile app version either.
Four known vulnerabilities remain to be addressed: synonym substitution, leet spelling variants, exfiltration instructions hidden in summaries, and non-English override instructions.
The layered design skeleton of this approach includes: separation of permissions and identity, external content untrusted by default, choosing denial when uncertain, and leaving a trace of all decisions.
All Facts (7)
The layered design skeleton of this approach includes: separation of permissions and identity, external content untrusted by default, choosing denial when uncertain, and leaving a trace of all decisions.
50%UnverifiedIn QuietOS's design, the Agent holds only identity while QuietOS holds the permissions; approvals expire, and there are verifiable logs.
50%UnverifiedQuietOS's proof log records every hold and decision with SHA-256 hashes in an append-only ledger, which can be audited afterwards.
50%UnverifiedA developer shared the QuietOS+Sentinel approach on Reddit, in which QuietOS handles outbound permission control and Sentinel handles inbound content review, with the final decision left to humans.
50%UnverifiedThe developer explicitly stated that there is no monthly fee, and there is not yet a mobile app version either.
50%UnverifiedFour known vulnerabilities remain to be addressed: synonym substitution, leet spelling variants, exfiltration instructions hidden in summaries, and non-English override instructions.
50%UnverifiedThe project has released its code publicly, published version 0.1.0 on PyPI, and passes 34/34 of its benchmark test suite.
50%