GuardAgent-OS
A zero-trust security orchestration framework for autonomous AI agents, implemented in Python. Through a three-layer control loop of identity, gateway, and sandboxed execution, it performs deterministic interception and verification before tool calls are committed.
Timeline (last 90 days)
GuardAgent-OS is an institutional-grade, zero-trust, Python-implemented agent security orchestration framework, designed and open-sourced by developer eratulyadav1988, divided into three independent modules: identity, gateway, and sandboxed execution.
GuardAgent-OS instantiates an isolated AgentMachineIdentity for each autonomous session, confines the session to a specific micro-scope, and sets strict session spending caps.
GuardAgent-OS uses Python decorators to intercept tool calls before the code runs, validating parameter types and authorization scopes, with permission and scope decisions being deterministic and not dependent on model judgment.
GuardAgent-OS's outbound gateway sanitizes payloads before data is rendered to end users or written to logging platforms, to prevent leakage of sensitive information (such as UK National Insurance numbers or backend authentication keys).
GuardAgent-OS's validation overhead per tool-loop execution is under 5ms.
Test data provided by the author of GuardAgent-OS shows a containment rate of 99.4% against common prompt injection attacks.
The 99.4% containment rate comes from the author's own testing and has not yet been independently verified by a third party.
GuardAgent-OS takes a lightweight route of Python decorators plus type validation; the advantage is that it is easy to embed into existing codebases, and the cost is that security capabilities are tightly bound to the way the code is organized.
All Facts (8)
GuardAgent-OS is an institutional-grade, zero-trust, Python-implemented agent security orchestration framework, designed and open-sourced by developer eratulyadav1988, divided into three independent modules: identity, gateway, and sandboxed execution.
50%UnverifiedGuardAgent-OS instantiates an isolated AgentMachineIdentity for each autonomous session, confines the session to a specific micro-scope, and sets strict session spending caps.
50%UnverifiedGuardAgent-OS uses Python decorators to intercept tool calls before the code runs, validating parameter types and authorization scopes, with permission and scope decisions being deterministic and not dependent on model judgment.
50%UnverifiedGuardAgent-OS's outbound gateway sanitizes payloads before data is rendered to end users or written to logging platforms, to prevent leakage of sensitive information (such as UK National Insurance numbers or backend authentication keys).
50%UnverifiedGuardAgent-OS's validation overhead per tool-loop execution is under 5ms.
50%UnverifiedTest data provided by the author of GuardAgent-OS shows a containment rate of 99.4% against common prompt injection attacks.
50%UnverifiedThe 99.4% containment rate comes from the author's own testing and has not yet been independently verified by a third party.
50%UnverifiedGuardAgent-OS takes a lightweight route of Python decorators plus type validation; the advantage is that it is easy to embed into existing codebases, and the cost is that security capabilities are tightly bound to the way the code is organized.
50%