A local AI agent can use whatever files, credentials, tools, and network routes its process can reach. The dependable way to limit that access is defense in depth: run it in an OS-enforced sandbox or VM as an unprivileged user, expose only the task workspace, enforce outbound network rules outside the agent, and keep sensitive credentials beyond its reach. Prompts and agent-level permissions help, but they are not hard security boundaries.
Start with the environment, not the prompt
Code written or invoked by an agent runs with the capabilities available to its process. If that process can read a file, use a credential, or connect to a destination, the agent may be able to do so as well. OpenAI’s agent security guidance recommends isolated compute and warns against sharing an environment across unrelated users or trust boundaries.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
MINISFORUM MS-02 Ultra Workstation Mini PC, Intel Core Ultra 9 285HX (24C/24T, up to 5.5GHz), PCIe... | $1,659.00 | Buy on Amazon |
| 2 |
|
GMKtec EVO-X2 AI Mini PC Ryzen Al Max+ 395 Superchip 128GB LPDDR5X 2TB SSD | $3,649.99 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
Treat each agent session as potentially able to exercise the permissions of the account and sandbox in which it runs. An instruction such as “do not read secrets” may guide the model, but cannot prevent code from accessing a readable file. Use policy enforced by the operating system, VM, container, or network layer for the outer boundary.
Choose an isolation approach
These approaches can be combined. Agent-native controls are useful inside an OS-enforced environment, rather than as a replacement for one.
#1 Best Overall
- High-Performance AI Processor：The MS-02 Ultra features an Intel Core Ultra 9 285HX (24C/24T, up to 5.5 GHz, 13 TOPS NPU), delivering fast and efficient performance for AI inference, algorithm development, and media workloads. A PCIe x16 expansion slot supports desktop-class GPU upgrades for advanced model training and accelerated computing tasks. It's ideal for creators, engineers, and teams handling intensive parallel workloads.
- 4 × M.2 PCIe 4.0 + 4 × DDR5 SODIMM slots:Four DDR5 SODIMM slots support up to 256 GB of memory, while ECC helps maintain data integrity in mission-critical environments. Four PCIe 4.0 M.2 slots support up to 24 TB of storage, supporting RAID 0/1/5/10, combining high-speed performance with data protection. It allows for the creation of independent scratch disks, media libraries, and project drives, providing high-throughput for production workflows.
- PCIe & USB 4.0 v2: Up to three PCIe slots can be equipped, including a dual-slot x16 GPU. The main slot supports PCIe 5.0, meeting the needs of high-bandwidth creative and computing workloads. USB 4.0 v2 (80Gbps) supports high-bandwidth external storage and displays.
- Ultra-fast Networking: Wi-Fi 7 further enhances wireless performance with next-generation speeds and low-latency stability. Intelligent bandwidth switching optimizes throughput in different network environments, ensuring optimal performance for enterprise or local networks. Dual 25GbE ports (providing up to approximately 3.125 GB/s bandwidth, about 25 times faster than traditional 1GbE), enabling seamless large-scale file transfers and parallel computing. 10GbE and 2.5GbE ports, with support for Intel vPro technology, ensure enterprise-grade remote management and deployment flexibility.
- Server-grade thermal architecture: Utilizing a dedicated CPU/GPU airflow design, equipped with a 6-pipe dual-fan cooler, it maintains stable performance even under sustained loads, delivering up to 140W Turbo power while maintaining a 100W TDP, and operating with noise levels as low as 36 dB. An integrated 350W power supply ensures stable and reliable output for demanding computing tasks and fully loaded extended configurations.
| Approach | Filesystem boundary | Network boundary | Credential handling | Host services and trade-offs |
|---|---|---|---|---|
| Agent-native permissions | Controls are product-specific and may limit tools or working directories; the OS still determines what the process can access. | An agent allowlist is not external enforcement. Pair it with a firewall or sandbox policy. | Keep secrets out of directories and environment variables the agent can read. | Usually simpler to enable, but behavior and flags vary by agent and version. |
| Container or devcontainer | Use mounts and OS permissions to expose only required paths. A container is not automatically a complete security boundary; configure it deliberately. | Set policy at the container or host network layer; do not assume egress is denied by default. | Avoid mounting host credential directories or passing broad secrets into the container. | Can fit development workflows; host access and local-service connectivity need explicit configuration. |
| VM-based sandbox | Provides a separate environment boundary; still limit files copied or mounted into it. | Enforce egress at the VM or surrounding network layer. | Keep host credentials outside the VM and provide only narrowly scoped credentials when necessary. | Stronger separation can add setup and compatibility overhead. |
| Managed or self-hosted sandbox | Follow the product’s mount and workspace controls, then verify effective access. | Use documented outbound controls and confirm their behavior for the deployed configuration. | OpenAI’s self-hosted sandbox guidance says to keep the application API key outside the sandbox. | Convenience and host-service access depend on the specific product; Docker documents explicit policy exceptions for some host-local services. |
OpenAI describes filesystem mounts and controlled access to external systems in its sandbox documentation. Anthropic recommends separate workspaces and environments where trust boundaries differ in its Claude Code sandboxing guidance. These are vendor-specific descriptions, not a guarantee that every container, VM, or sandbox has identical protections.
Set up a restricted environment
- Choose the boundary. Run an agent that executes generated code in a dedicated VM, container sandbox, or other isolated environment. Keep unrelated projects and sensitive host data out of it.
- Expose a narrow workspace. Mount or share only the repository or task directory and any explicitly needed scratch path. Do not make the whole home directory available by default. Add paths only when a task requires them. For example, Claude Code’s CLI reference documents
--add-dirfor additional working directories; this is a product-specific convenience, not a substitute for OS permissions. See Anthropic’s CLI reference. - Use a low-privilege identity. Run the process under a dedicated unprivileged account or sandbox identity. Limit writes to the workspace and specific scratch locations; do not grant elevated host privileges for convenience. Anthropic’s Claude Code security guidance describes project-scoped writes and suggests devcontainers as an added layer.
- Deny network access by default. At the VM, sandbox, or firewall layer, allow only the inference endpoint and the external services required by enabled tools. If traffic must go through an HTTP proxy, also prevent direct connections that bypass it. OpenAI notes that software that ignores environment proxy settings may not honor proxy-based restrictions; see its Windows security engineering article.
- Keep credentials separate. Do not mount SSH directories, cloud credential files, password stores, or production secrets into the agent workspace. If a task requires access, prefer a broker or temporary credential limited to that task and scope. OpenAI’s self-hosted sandbox guidance advises keeping the application API key outside the sandbox.
- Test the effective policy. Confirm the agent can read and write the intended paths, is denied access to out-of-scope paths, and cannot reach blocked destinations. Repeat checks after changing the provider, CLI version, plugins, or MCP servers, since those changes can alter what the process needs or can do.
Allow network access deliberately
Build the network allowlist from the agent’s actual configuration: the inference provider, plus only the services used by enabled tools. Do not copy an endpoint list from a different product or deployment and assume it is complete. Anthropic’s proxy guidance lists api.anthropic.com, statsig.anthropic.com, and sentry.io for the setup it documents; that list is not a universal allowlist for every Claude Code installation or tool configuration. See Anthropic’s network configuration documentation.
Local inference can still require network access if the model server runs outside the sandbox. Docker’s local-model walkthrough shows a sandbox isolated from the network until policy explicitly permits access to a host-local model endpoint. Treat that exception narrowly: allow the required host service and port rather than opening general access to the host or internet. The example describes Docker’s configuration, not the default behavior of all sandbox products.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #2
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Use agent controls as an inner layer
Agent controls can reduce accidental access and make a session easier to manage. For example, Claude Code documents directory additions, tool allow/deny flags, and a --dangerously-skip-permissions option in its CLI reference. Permission flags govern that product’s behavior; do not treat them as a replacement for OS-level file permissions or externally enforced network policy.
Codex behavior also depends on the installed version and configuration. An older OpenAI Help Center description says Full Auto is sandboxed, network-disabled, and scoped to the current directory. Because that description may not match later releases, verify the behavior against the official documentation for the version in use rather than relying on a mode name or older summary: OpenAI Help Center.
What to verify before relying on the setup
- Filesystem: Can the agent read and modify only the intended workspace and scratch locations? Check that home-directory files and mounted secrets are inaccessible.
- Network: Can it reach the required inference and tool endpoints, while direct connections to unapproved destinations remain blocked?
- Credentials: Are any tokens, keys, or credential helpers available to the process broader than the task requires?
- Host access: If a local model or other host service is required, is only that service reachable through an explicit exception?
- Change control: Do you repeat the checks when changing the agent, provider, plugins, MCP servers, sandbox settings, or network policy?
Product controls and defaults can change across versions, platforms, and deployments. Consult the current vendor documentation for the exact release in use, and validate the policy in that environment rather than assuming a label such as “sandboxed” means every file and network route is restricted.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Repair Windows errors before they cause bigger problems3Scan for outdated or missing drivers - takes under a minute
