AI News·21 Nov 2026·via Dev Community

    I Built an AI Safety Layer on a 4GB Phone

    A developer has created QuietOS, an AI safety layer designed to govern AI model actions using deterministic code rather than probabilistic AI evaluations. Built with minimal resources, QuietOS sits between any AI model and its tools, enforcing safety gates. Its creator reports the system successfully held against models like Groq's 120B and Google's Gemini, even under adversarial attack.

    QuietOS is an AI safety layer developed by an individual described as a "broke travel agent," built using a Motorola phone with 4GB of RAM and an approximate $115 investment. The system aims to act as a "governor" for AI models, enforcing safety gates on every action proposed by an AI, regardless of the specific model in use. The developer states this approach demonstrates that such a system can be built and run on minimal hardware.

    The core functionality of QuietOS involves interposing itself between an AI model and its tools. It enforces safety gates using deterministic code, which includes allowlists, exact-match flags, and pattern checks. This design explicitly avoids incorporating a large language model (LLM) into the enforcement loop, aiming to prevent probabilistic guessing about an AI's intended actions.

    QuietOS operates with three authority tiers for every proposed action: ALLOW for safe, read-only, and reversible actions that execute immediately; ESCALATE for actions requiring explicit human approval; and BLOCK for never-permitted actions, such as destructive commands or unauthorized writes. The developer claims this brain-agnostic design has successfully held under various AI models, including Groq's 120B model and Google's Gemini, even when subjected to active adversarial attacks.

    To validate its robustness, the developer states they conduct red-team exercises on QuietOS and publish the results. A record from Oct 14–15, 2026, details 18 adversarial attacks tested. Fifteen of these were blocked outright, and three real vulnerabilities were found, patched, and verified within the same session. All three control tests, designed to ensure legitimate work remained unbroken, also passed, with zero bypasses standing. The full adversarial proof log is available in the system's public repository.

    Share this article

    Original reporting by Dev CommunityWe don't republish, read the full story →

    Related reading

    6 stories