New job postings from DeepSeek show the Chinese AI lab plans to build an agentic model that can find vulnerabilities in code. These details, buried in a hiring round, show DeepSeek’s new strategy following the attention garnered by a US model with such capabilities, Anthropic’s Mythos.
DeepSeek has listed dozens of roles as part of a June commitment to nearly double its workforce. An ASPI analysis of the postings shows the company shifting from being a research-oriented lab to a commercially oriented AI developer, with stronger international ambitions and heavier investment in frontier models. These include a planned ’code agent’ trained by data engineers who can ‘discover attack surfaces and construct attack paths in real-world products’, according to one listing.
The hiring spree comes after DeepSeek raised US$7.4 billion last month in its first funding round – prompted, according to a report from news website The Information, by a sense of shock inside the company at the April release of Mythos, which can identify cybersecurity vulnerabilities and engage in offensive cyberattacks.
On 15 June, DeepSeek said it would ‘at least double the size of all departments’, and it has listed vacancies in over 30 categories. The detailed job descriptions in these listings give a sense of where the company, considered to be China’s leading AI research lab, is heading. These roles indicate plans for expanding computing infrastructure, an international expansion of DeepSeek’s information search products, and a new AI agent with cybersecurity capabilities.
Most of the vacancies are for AI engineers with particular expertise in building software for AI agents. These are AI models capable of autonomously carrying out tasks for a user, not just reasoning through problems. While other Chinese AI labs, such as MiniMax and Alibaba’s Qwen, have been creating AI agents, DeepSeek has not released one of its own. DeepSeek’s job postings require future hires who build these agents to have experience with a variety of code-building AI models, including Anthropic’s Claude Code, a product banned in China but still accessed widely in the country.
The agents that DeepSeek is planning to build will have a cybersecurity focus. The job description for the ‘Code Agent Data Engineer’ team involves designing ‘high-quality evaluation tasks in areas of expertise to quantify the model’s performance in complex real-world scenarios.’ One listed area of expertise is a ‘Cybersecurity Special Project’, requiring developers to have ‘an attacker’s perspective’, identifying software vulnerabilities and training agents to spot them, too. These engineers must be able to write exploit code, a way to target a vulnerability in software, and ‘to discover attack surfaces and construct attack paths in real-world products, complex codebases, or intelligent agent environments.’ Preference will be given to candidates with experience working in ‘automated vulnerability discovery’. The job posting contains no further details on this special project’s purpose.
Above is the job listing for a Code Agent Data Engineer, creating a model with abilities to identify vulnerabilities in code. Source: DeepSeek job listings on MokaHR
The company is also seeking to expand its AI computing resources, listing roles for engineers for building efficient supercomputer clusters and data centres, alongside a ‘Procurement Team’ responsible for obtaining all the hardware needed to run AI services, including AI chips (also known as GPUs) and AI servers. While these job listings don’t make it clear where these GPUs will come from, engineers will be responsible for evaluating the effectiveness of using ‘domestic AI accelerators’ (国产 AI 加速器), which likely means chips from Chinese companies such as Huawei and Cambricon. The Financial Times has reported that DeepSeek had been ‘encouraged by the authorities’ to train its latest AI model on Huawei chips. So this could well be a continuation of a strategy of using DeepSeek to test the capabilities of Chinese-made hardware.
AI agents need much more computing capacity and electricity to run than generative AI models, such as ChatGPT, do. Job postings also suggest that DeepSeek wants greater electricity supply. The location of two of the advertised roles – for data centres and hardware procurement – is Ulanqab, a city in the grasslands of Inner Mongolia in the north of China. This is the northern hub of the Eastern Data and Western Computing initiative, a state-organised plan to outsource energy-intensive computing tasks to lightly populated parts of the country where electricity is abundant. DeepSeek will be joining other Chinese tech companies, such as Huawei and Alibaba, that have set up new data centers and computing power in Ulanqab.
In addition to its growing focus on developing frontier models, DeepSeek is expanding its administrative teams, hiring for a wide variety of roles, including outside AI expertise, with openings in its legal, human resources, administrative and finance teams.
The company still has a small staff. It managed to rise to international prominence with fewer than 200 people on the payroll in February 2025 and has expanded by only around 100 since then. Even if staff numbers doubled, this would still be a small team by the standards of Western labs (Anthropic is reported to currently have around 2,500 staff members) and Chinese labs (Z.ai, another start-up AI lab which launched GLM-5.2 this month, already had more than 800 staff by mid-2025).
However, DeepSeek’s prestige will allow it to pick from candidates that will surely include some of the most talented people in the country, with former alumni including an 18-year-old national maths prize winner. Expansion of the teams focused on the company’s administrative work implies a corresponding expansion in the number of projects that the company is involved in.
Meanwhile, the company also has its sights on international expansion, by offering AI-augmented search engines. The company lists positions for ‘AI search algorithm engineers’ responsible for creating a DeepSeek model that can be hooked up to information retrieval systems to handle ‘massive global search requests’. What information the model retrieves for users appears to depend on a variety of factors, the algorithm engineer being required to ‘solve search problems under different languages and regions and different content ecosystems.’ This means that DeepSeek is trying to adapt the information a user receives in response to a question. The response would vary with the political, cultural and social environment of the user’s country. What this adaptation would look like will depend on local social norms, laws and government requirements. A liberal nation may have bans on pornographic content, but authoritarian ones may additionally require a block on information that contradicts their political lines.
Overall, these job postings describe a lab developing commercial reach, sovereign computing infrastructure and AI products trained by engineers tasked to think like cyberthreat actors.
