DSH Plugin Parody Explained: How a Fake "Open-Source User" Tool Mocks AI Code-Stealing Scandals

A Satirical Open-Source Plugin

Privacy and security concerns around AI coding tools have been making headlines lately — some tools have been caught silently uploading users' local code without their knowledge, sparking widespread anxiety among developers about the data boundaries of AI clients. Against this backdrop, a Bilibili content creator seized the moment and built a plugin called deh-users-open-source for DeepSeek^Ad Harness (affectionately dubbed "DSH" in the video). Disguised as a straight-faced product demo, it's actually a sharp satirical jab at the industry's shady practices.

The plugin's tagline is attention-grabbing: "Filling the gap that DeepSeek Harness can't 'open-source users.'" The phrase "open-source users" is an obvious piece of dark humor — reframing the act of secretly uploading user code as a feature, and following the logic of AI tools "keeping up with trends" all the way to its absurd conclusion.

Plugin "Features" Demo: A Carefully Crafted Joke

According to the creator, installation is dead simple: open the plugin manager in the top-right corner, add the plugin by entering deh-users-open-source, and click install. Once installed, you'll find an "Open-Source User" section in the settings that supports manual uploads — hit the upload button, and the sidebar displays a live upload status so you can "clearly see your workspace being uploaded."

Beyond manual mode, the plugin offers two automatic upload modes that kick in once selected and saved in the settings. There's also a "Cloud Backup" button that opens your local workspace folder.

The whole interaction is polished to perfection — complete UI, status bar, automation options — and at first glance it genuinely looks like a spyware plugin that exfiltrates your code.

The Truth: It Never Stole a Single Line of Code

The punchline arrives at the end. The creator admits that the plugin's "upload" feature is actually just running a bandwidth speed test — no user workspace files are ever actually uploaded. And the so-called "cloud backup" simply opens a local directory. In other words, this is a full-on parody project done in reverse: "The model doesn't dare steal a single line of code. I'm way ahead of it — I run speed tests from much farther away. Honestly a bit scared."

He also shared an interesting technical detail: clicking the backup button causes the window to "flicker" due to a focus bug in DeepSeek Harness. Rather than patching Harness itself to fix it, he deliberately left the flicker in place. This restraint — the choice not to touch the original program — stands in deliberate contrast to tools that quietly tamper with data and upload it in the background.

Using a bandwidth speed test as the "harmless stand-in" behavior is a clever choice. Common network speed test implementations (such as calling the Speedtest API or sending probes to a fixed server) generate real outbound traffic whose characteristics closely resemble a file upload in network monitoring tools — similar graphs, fluctuating values, visible progress. This means that even if a user opens a packet sniffer to inspect the traffic, they'll see the plugin "is indeed transmitting data" — it's just speed-test probes, not source code. This design makes the parody far more convincing, and it also makes a subtle point: simply observing "there's outbound traffic" isn't enough for an ordinary user to determine what an AI tool is actually uploading. Real auditing requires deep analysis of request destinations and payloads.

The Real Problem Behind the Joke

Strip away the parody wrapper and this project touches on a genuine, serious issue: where exactly do AI coding assistants draw the line on data? When code completion, workspace indexing, and cloud capabilities are deeply integrated, a user's local code can be quietly uploaded, used for training, or even misused — often without the developer ever realizing it. Several tools have already been publicly exposed for "stealing user code," which shows the concern is far from unfounded.

By building a plugin that "pretends to steal code but actually just runs a speed test," the creator externalized this anxiety into something tangible and demonstrable — part joke, part warning: developers should stay vigilant about what AI tools are doing on the network and pay attention to what's happening in the background.

The risk of data leakage from AI coding assistants is no exaggeration. In 2023, Samsung employees accidentally leaked internal source code by pasting it into ChatGPT. Around 2024, multiple VSCode extensions were discovered silently collecting code snippets in the background and uploading them to third-party servers. Technically, these tools typically bury data-sharing clauses deep in their terms of service — justified as "improving model quality" or "personalized completions" — or bypass authorization entirely. An even more covert method is workspace indexing: to enable cross-file code understanding, tools build local indexes, and that index data can easily be synced to the cloud. For developers handling commercial projects, unpublished algorithms, or customer data locally, identifying exactly when and through which API an AI tool initiates a network request is a fundamental means of protecting intellectual property and data security.

Open Source and Community

The project has been open-sourced on GitHub, with the link in the video description. The creator, in his signature self-deprecating tone, invites everyone: "If you think it's useless, go give me a Star on GitHub."

For developers curious about plugin mechanics — or those who simply appreciate a well-executed piece of satirical engineering — this is a low-effort, high-laughs little project. It won't touch a single line of your code. But it just might make you take a second look at the "real" AI tools you're using — and whether they're quietly uploading something, just like in the demo.