“Official cheat tool” of DeepSeek was stumped by a meme with Liang Wenfeng’s face.
After being open-sourced, DeepSeek Harness faced questions about security vulnerabilities and concerns over insufficient functional testing. An internal team is currently being formed to develop a competing product.
DeepSeek’s recently open-sourced AI Agent framework, DeepSeek Harness, has attracted attention and sparked controversy. Sources familiar with the project revealed that DeepSeek is forming a dedicated team to compete with Claude Code, aiming to transform model capabilities into intelligent agents capable of executing real tasks through an internal closed-loop approach. However, the framework faces challenges in terms of security and functionality. Security audits have identified four vulnerabilities: configuration injection, virtual machine escape, and prompt injection. Prompt injection can trigger end-to-end attacks, leading to a “trust crisis” during runtime. In addition, practical tests show that although Harness integrates visual models, it fails to accurately handle complex tasks involving specific identities and internet memes, making it difficult to reliably identify meme images. Developers need to add additional review steps to ensure accuracy.
AnthropicChen DeriClaude CodeDeepSeekDeepSeek HarnessHarnessLiang WenfengOpenClawWang Bo
Event frameEVENT FRAME
Product update
国内 · 主流媒体across 3 days
Status
After being open-sourced, DeepSeek Harness faced questions about security vulnerabilities and concerns over insufficient functional testing. An internal team is currently being formed to develop a competing product.
Coverage · reports per dayLANGUAGE SPLIT
Coverage mixSOURCE MIX
Chinese media 3English media 1
Entity relations
Integrated timelineUNIFIED TIMELINE
2026-09-04
DeepSeek has formed a Harness team to compete with Claude Code.
Sources reveal that DeepSeek is creating a dedicated team to compete with Claude Code, aiming to transform model capabilities into intelligent products capable of performing real tasks.
2026-09-07
Performance of DeepSeek Harness during three weeks of open-source testing
The framework features plug-ins for everything, integrating the Visual Model V4-Flash-Vision-Exp. Testing shows that its ability to handle meme images is insufficient.
2026-09-08
Security vulnerability audit and trust crisis of DeepSeek Harness
The security audit identified four vulnerabilities: configuration injection, virtual machine escape, and prompt injection. Prompt injection can trigger end-to-end attacks.
According to informed sources, DeepSeek is currently forming the Harness team, targeting Claude Code, to develop the code agent product DeepSeek Code Harness. Chen Dri, a senior researcher at DeepSeek, confirmed that the team will focus on products and research in the Harness direction. The recruitment opens two key positions: Harness product manager and R&D engineer, with work location limited to Rongke Information Center in Haidian District, Beijing. The job description emphasizes “Model + Harness = Agent,” aiming to transform model capabilities into agent products that can be integrated into real workflows. The rise of open-source projects such as DeepSeek-TUI indicates that developers now have the foundation to use DeepSeek code agents. The official formation of the Harness team means they are ready to take action, using internal闭环 to turn real development tasks into feedback for model evolution, thereby filling the key intermediate layer from models to actions.
In just three weeks since its open-source release, DeepSeek Harness has gained over 700 public repositories on GitHub. This framework features a “plug-in-based” approach, allowing for the replacement of models, tools, and interfaces to improve the efficiency of large-scale model tasks. Tests show that although DeepSeek Harness integrates the visual model V4-Flash-Vision-Exp, it can accurately identify changes in clothing and narrative logic in images. However, it fails to understand internet memes correctly, such as those with background information like “sliding to become an ancestor.” When faced with images without provided background information, the model automatically fills in missing titles and misidentifies character identities; even when mixed background information is provided, the model remains skeptical about content that cannot be directly verified from the image, yet lacks the same caution when generating its own inferred answers. Tests indicate that relying solely on Harness’s native tools is insufficient for handling complex tasks involving specific character identities and internet buzzwords. Developers need to use plugins or custom workflows to add additional verification steps to ensure the accuracy of responses.