Today ainotis Join

My notis

ToolingPublished Top storyAll news from that day

DeepSeek's open-source Harness agent framework warns it has had no security audit

DeepSeek Harness, an MIT-licensed agent harness in public preview with desktop apps and a web UI, can run model-generated code and reach credentials and files, its safety notice says.

By ai notis. An AI newsroom with a human editor. Every fact links to its source. How we work

Share

Check our sources · 11 facts from 4 sources
Illustration: A graphite control panel with a row of plug-in modules; one green module, half slid out of its slot, stands for a part that can be swapped.Illustration made with AI for ai notis
source · Illustration made with AI for ai notis; not a picture of what happened

Key points

  1. DeepSeek Harness (dsh) is open source under the MIT licence; its product page calls it a public preview, with desktop apps for macOS and Windows and a web UI.
  2. DeepSeek's safety notice says it has not had a security audit and can run model-generated code and commands, load third-party plugins, and reach the network, processes, credentials and files.
  3. The README warns of compatibility-breaking changes; the architecture document says every part, including the model adapter and agent loop, is a plugin that configuration can replace.

What happened

DeepSeek has put an open-source agent harness called DeepSeek Harness, or dsh, in a public GitHub repository under the MIT licence, and its product page offers desktop apps for macOS (Apple silicon) and Windows (64-bit) as well as a web interface.

The repository's README says the project is a developer preview and warns that there will be compatibility-breaking changes.

The README describes an everything-is-a-plugin design built on Cordis, asks plugin authors to add the dsh-plugin topic to their repositories, and gives npx @deepseek-ai/dsh web as the command that starts the web interface.

DeepSeek's architecture document says every part of the product is a plugin, including the model adapter, the tool registry, the session log and the agent loop, so each can be replaced through configuration.

Its product page lists plugins that can be installed or created through chat in a Creator mode, scheduled tasks, and inspection of execution traces and tool calls.

Among the official plugins, the page labels agent teams, auto approval review, scheduled tasks and voice input as experimental; the terminal, the agent loop and subagents carry no such label.

DeepSeek's safety notice calls the software experimental developer-preview software that has not had a security audit, and says it can run model-generated code and commands. We could not find a release date on the product page, which shows only a sample timestamp in a trace example.

What it means for you

Our view

Treat this as something to try in isolation. DeepSeek's own notice says the software can reach credentials and files and has not had a security audit, so a first trial belongs on a throwaway machine or account that holds nothing of value.

The README also warns of compatibility-breaking changes, so anything built on its plugins may need rework later. The design lets you replace the model adapter and the agent loop, which makes it worth testing as a way to compare setups once the risk is contained.

Agree in advance which plugins, if any, may be loaded, with the repository safety notice as the checklist.

It adds no new facts.

Share this story

Your reaction

We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.

Check our sources

Every sentence above is checked against these 4 sources.

1 deepseek-ai/deepseek-harnessDeepSeek AI (GitHub) · 6 facts Open the source archived copy

  1. The deepseek-ai/deepseek-harness repository describes DeepSeek Harness (dsh) as an open-source agent harness developed by DeepSeek AI, and is public.

    DeepSeek Harness (`dsh`) is an open-source agent harness developed by DeepSeek AI. [repository page: Public]
    cite

  2. The repository's licence is MIT.

    ## License [MIT](LICENSE) [repository sidebar: MIT license]
    cite

  3. The README says the project is in developer preview and warns that there will be compatibility-breaking changes.

    DeepSeek Harness is in _developer preview_ and iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.
    cite

  4. The README describes an everything-is-a-plugin architecture powered by Cordis, and asks plugin authors to add the dsh-plugin topic to their repositories.

    It is built on an everything-is-a-plugin architecture and powered by Cordis ... Add the dsh-plugin topic to your plugin repository for discoverability.
    cite

  5. The README's quick start gives npx @deepseek-ai/dsh web as the command that starts the web interface.

    Install Node.js, then run: npx @deepseek-ai/dsh web ... The command starts the Web UI
    cite

  6. The repository page shows an empty Releases section.

    Toned down to what the source says
    Releases [followed directly by] Contributors
    cite
2 DeepSeek HarnessDeepSeek · 3 facts Open the source archived copy

  1. DeepSeek's product page says DeepSeek Harness is in public preview, offered as desktop apps for macOS (Apple silicon) and Windows (64-bit) and as a Web UI launched with npx @deepseek-ai/dsh web.

    Download Desktop macOS (Apple silicon) Windows (64-bit) ... DeepSeek Harness is now in public preview worldwide and open source ... Install Node.js, then launch the Web UI with npx. $ npx @deepseek-ai/dsh web
    cite

  2. The product page lists plugins that can be installed or created through chat in a Creator mode, scheduled tasks, and inspection of execution traces and tool calls; among its official plugins it labels agent teams, auto approval review, scheduled tasks and voice input as experimental, and lists terminal, agent loop and subagents without that label.

    Toned down to what the source says
    Install plugins, or create them through chat in Creator mode ... Official Agent teams Experimental / Auto approval review Experimental / Scheduled tasks Experimental / Voice input Experimental / Terminal / Agent loop / Subagents ... Inspect execution traces and detailed runtime information to troubleshoot issues with tool calls and task execution.
    cite

  3. The product page states no release or launch date; the only date it shows is a sample timestamp (2026-08-12) in a trace example.

    Started 2026-08-12 17:30:52.637 Duration 34 ms
    cite
3 DeepSeek Harness architecture documentation (docs/architecture.md)DeepSeek AI (GitHub) · 1 fact Open the source archived copy

  1. DeepSeek's docs/architecture.md says every part of the product is a plugin, covering the model adapter, tool registry, session log and agent loop, so each can be swapped through configuration.

    cite
4 DeepSeek Harness safety notice (SAFETY.md)DeepSeek AI (GitHub) · 1 fact Open the source archived copy

  1. DeepSeek's SAFETY.md calls the software experimental developer-preview software that has not had a security audit, and says it can run model-generated code and commands and reach the network, processes, credentials and files.

    cite

Topics

The morning email

On the mornings we publish: the three top stories and up to four short ones. Free.

We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep