Tenuo shipped an integration with NVIDIA OpenShell that layers task-scoped authority on top of OpenShell's sandboxing. OpenShell controls what an agent can reach, while Tenuo limits what the agent can do per task — e.g. time-limited, repo-scoped GitHub push permission. The integration targets agent-security practitioners building sandboxed agent runtimes.
A developer named Rajan built a Temporal agent using Tenuo and planted a prompt injection in robots.txt instructing it to read .env. The agent attempts the injection but the call is rejected on the worker before the activity runs, after which it completes only its authorized tasks. The demo illustrates combining Temporal workflows with scoped authority for AI agent safety.
A developer named Rajan built a Temporal agent using Tenuo and planted a prompt injection in robots.txt instructing it to read .env. The agent attempts the injection but the call is rejected on the worker before the activity runs, after which it completes only its authorized tasks. The demo illustrates combining Temporal workflows with scoped authority for AI agent safety.
A Tenuo developer argues the SalesBleed incident shows AI agents need task-scoped permissions — a 'valet key' model granting only what the current task requires, rather than full role-based access. Tenuo is an Apache 2.0 open-source project whose AAT Internet-Draft was recently cited by NIST in its work on agentic AI identity and authorization.
A Tenuo developer argues the SalesBleed incident shows AI agents need task-scoped permissions — a 'valet key' model granting only what the current task requires, rather than full role-based access. Tenuo is an Apache 2.0 open-source project whose AAT Internet-Draft was recently cited by NIST in its work on agentic AI identity and authorization.
Sorami Consulting published an independent evaluation of NVIDIA OpenShell v0.1.2, an open-source sandbox runtime released 28 September under the Open Agent Safety Platform. Across 123 trials with a local qwen3:8b agent, every documented default control held, including blocking a malicious setup script that leaked a secret 10/10 without the sandbox. The main failures came from operator settings: automatic approval granted outbound access to new hosts 12/12 times.
Sorami Consulting published an independent evaluation of NVIDIA OpenShell v0.1.2, an open-source sandbox runtime released 28 September under the Open Agent Safety Platform. Across 123 trials with a local qwen3:8b agent, every documented default control held, including blocking a malicious setup script that leaked a secret 10/10 without the sandbox. The main failures came from operator settings: automatic approval granted outbound access to new hosts 12/12 times.
A developer building internal AI agents asks how to design access control for agents that call APIs and touch customer data, arguing binary allow/deny models built for human roles don't fit. The post frames agent access as task-scoped rather than role-based, with minimum privileges earned per task. It seeks practitioner input on whether scoping should happen at the agent level or the individual tool-call level.
A developer building internal AI agents asks how to design access control for agents that call APIs and touch customer data, arguing binary allow/deny models built for human roles don't fit. The post frames agent access as task-scoped rather than role-based, with minimum privileges earned per task. It seeks practitioner input on whether scoping should happen at the agent level or the individual tool-call level.
AgentZ is an open-source governance layer that wraps any agent with sandboxing, credential isolation, and workspace boundaries. It targets teams running agents via open frameworks who currently manage permissions ad hoc with spreadsheets and trust. LLM agents are the runtime being governed, not provided by AgentZ itself.
AgentZ is an open-source governance layer that wraps any agent with sandboxing, credential isolation, and workspace boundaries. It targets teams running agents via open frameworks who currently manage permissions ad hoc with spreadsheets and trust. LLM agents are the runtime being governed, not provided by AgentZ itself.