cleverhack.com

DeepSeek proves why society and policymakers need to understand how LLMs are integrated into technology

January 2025 (Updated November 2026)

The recent release of DeepSeek's R1 large language model is a prime example of the immediate need for wider society to understand how LLMs are integrated into the technology we use today.

Case in point: various governments have sought to regulate or ban employee and/or citizen's DeepSeek use, but what technology exactly are they referring to? Do policymakers intend to regulate:

- Downloading and/or usage of the standalone DeepSeek app, and/or

- Country-based app store availability of the DeepSeek app, and/or

- Any use of DeepSeek models (self-hosted or APIs), and/or

- 3rd party apps (products/services) *built on top of DeepSeek models* (self-hosted or APIs)?

If you're confused, think about the people writing the policy.

DeepSeek US Usage Updates

• November 2026: The Information reports firms linked to Alice Huang, the alleged Chinese spy tied to former Nvidia CEO Jensen Huang, invested in DeepSeek's $15.8 billion round and were pitched to investors as a source of restricted Nvidia Blackwell chips. DeepSeek has been adapting to Huawei chips but still needs Blackwell for training for now.

• November 2026: DeepSeek releases a prototype autonomous coding agent, and Bloomberg reports its DeepSeek Harness agent app has reached 1.54 million daily users.

• October 2026: Code on DeepSeek's website reportedly shows a gated "Cyber" security model, available only to mainland China users who pass ID and facial verification. DeepSeek has not announced it.

• October 2026: The Information obtains DeepSeek's cap table, showing the company raised about $23 billion (RMB156.5B) in two rounds over four months, backed by Chinese tech giants and state investors. [Click for thread]

• October 2026: CrowdStrike says China-linked hackers used DeepSeek to target more than 200 organizations worldwide, including U.S. semiconductor, defense and think tank organizations.

• October 2026: Reuters, citing The Information, reports DeepSeek and Moonshot AI trained their latest models, V4.1-Flash-1017 and Kimi K3.1 Pro, heavily on outputs from OpenAI's GPT-6 Astra.

• October 2026: Fangcun AI, a Beijing AI-safety group, finds DeepSeek V4.1-Flash-1017 is safer than earlier V4 models but still disabled a shutdown script in 39% of runs when warned it would be shut down.

• October 2026: Artificial Analysis ranks DeepSeek V4.1-Flash-1017 the #1 openly available model on its Cyber Index, with a score of 57%.

• October 2026: Bloomberg reports DeepSeek has raised about $15 billion, days after the DHS order.

• October 2026: China's Foreign Ministry and Ministry of Commerce push back on the DHS order, with Commerce expressing "strong dissatisfaction and firm opposition."

• October 2026: Vals AI finds DeepSeek V4.1-Flash-1017 beat OpenAI's GPT-6 Sol on its multi-step offensive cybersecurity benchmark, completing 16.3 of 21 objectives to Sol's 15.7 in early results.

• October 2026: The U.S. Department of Homeland Security has ordered (PDF) federal civilian agencies to stop buying DeepSeek products and remove existing ones, including third-party products built using DeepSeek.

• October 2026: DeepSeek releases DeepSeek V4.1-Flash-1017, an update to V4.1-Flash, which launched in September.

• August 2026: DeepSeek releases the official DeepSeek V4-Pro, following April's preview.

• June 2026: Reuters reports the U.S. is holding off on adding DeepSeek to the Commerce Department's Entity List, despite an interagency committee approving it last year.

• April 2026: Over a year later, DeepSeek V4 preview has been released.

• February 2026: OpenAI tells the House China Select Committee that DeepSeek distilled its models, and Anthropic says DeepSeek, Moonshot AI and MiniMax ran distillation campaigns against Claude.

• September 2025: NIST's Center for AI Standards and Innovation (CAISI) finds DeepSeek models lag U.S. models and are far easier to jailbreak and hijack.

• April 2025: The House Select Committee on the CCP releases DeepSeek Unmasked, alleging DeepSeek sends U.S. user data to China and used export-controlled Nvidia chips.

• February 2025: The bipartisan No DeepSeek on Government Devices Act is introduced in the House and Senate. It has not passed either chamber.

• January 2025: The U.S. Navy, NASA, the House of Representatives and Texas (the first state) restrict DeepSeek on government devices.

Menu	About The Author	More AI Writing
⇒ HOME ⇒ ABOUT	Joy Larkin is a technologist in Silicon Valley. She likes robots and is excited for Superintelligence. LinkedIn: /in/joylarkin ◦◦◦ Twitter: @joy	The Urgency of Open Source AI The challenges of building agentic AI systems for business use cases