Major AI Companies Begin Large-Scale Recruitment of Weapons Experts

AI companies Anthropic and OpenAI have begun recruiting chemical weapons and explosives defense specialists to prevent their AI models from being exploited for weapons of mass destruction (WMD) manufacturing. According to a mid-March BBC report, Anthropic posted a job listing on LinkedIn seeking chemical weapons and high-explosives experts for "catastrophic misuse prevention." The position requires at least 5 years of experience in chemical weapons or explosives defense and knowledge of radiological dispersal devices (dirty bombs).

OpenAI is similarly recruiting researchers for "biological and chemical threats," offering salaries up to $455,000 (approximately 530 million won)—nearly double Anthropic's salary range. Both companies have effectively acknowledged through these job postings that their models have reached a level where they can provide "meaningful assistance" to individuals with malicious intent.

Why WMD Specialists Now?

This recruitment push goes beyond simple preemptive safety measures. It reflects the industry's own assessment that AI model capabilities have crossed a certain threshold. Anthropic's Claude and OpenAI's GPT series already demonstrate expert-level outputs in generating complex chemical reaction formulas, explaining experimental procedures, and proposing substance synthesis pathways.

The timing of these job postings is particularly significant. It coincides with reports that the U.S. Department of Defense is deploying AI tools at an unprecedented scale during the three-week-old armed conflict in Iran. As military applications of AI technology accelerate, the risk of exploitation by non-state actors or hostile forces is becoming a reality.

Dr. Stephanie Hare, a technology researcher and co-host of BBC's AI Decoded program, expressed concern: "Is it safe for AI systems to handle sensitive information about chemicals, explosives, and radiological weapons including dirty bombs?" She pointed out that "there are no international treaties or regulations governing this work, and everything is happening behind closed doors."

What Has Changed?

AI companies' safety verification strategies are undergoing qualitative transformation. While past approaches centered on 'red teams' composed of computer scientists and ethicists, the current phase involves directly recruiting specialists with hands-on experience in actual weapons systems.

OpenAI's exceptional maximum salary of $455,000 indicates how seriously the AI industry is taking this issue. This represents 1.5 to 2 times the salary of typical AI researchers or software engineers, suggesting a market premium is forming for this specialized expertise.

Conflicts Over Defense Collaboration

Both companies have shown conflicting internal positions regarding collaboration with the Department of Defense. OpenAI changed its defense-use policy in 2024 to open possibilities for military cooperation, but faced pushback from some employees. While Anthropic has maintained a relatively cautious stance, an indirect connection has formed through its cloud infrastructure provider AWS's participation in Defense Department projects.

The specific sources of AI tools currently used by U.S. military forces in the Iran conflict have not been disclosed, but industry reports suggest that large language model (LLM)-based intelligence analysis systems are being utilized in operational planning. This reveals the reality that AI companies cannot fully control the direct and indirect military application pathways of their technology.

Technical Dilemma: Growing Risks with Information Injection

Paradoxically, the safety enhancement strategy through specialist recruitment contains new risks. In the process of building blocking mechanisms by injecting weapons manufacturing knowledge into AI models, this information may leave traces in the model's training data or parameters in some form.

Dr. Hare's point about "the risk inherent in information processing itself" refers to this issue. There is a fundamental difference between a model "knowing but not answering" chemical weapons manufacturing procedures and "not knowing in the first place." In the former case, the possibility of circumvention through adversarial prompt engineering or jailbreak techniques remains constant.

The international community currently lacks treaties or standards governing AI systems' handling of sensitive information. Existing arms control frameworks like the Chemical Weapons Convention (CWC) or Nuclear Non-Proliferation Treaty (NPT) are not suited to address the dual-use nature of AI technology. Each company's self-established 'Responsible AI' principles serve as the only guidelines.

[AI Analysis] Future Outlook and Implications

Industry-Wide Standardization Movements Expected
Major AI companies are likely to jointly establish 'Hazardous Materials Information Processing Guidelines' within the next 6-12 months. Given the existence of industry consortia like Partnership on AI and Frontier Model Forum, as cases of WMD specialist recruitment accumulate, the industry will likely advance to a best practices sharing phase.

Government Intervention to Fill Regulatory Gaps
The U.S. National Institute of Standards and Technology (NIST)'s AI Risk Management Framework (AI RMF) may specify weapons exploitation scenarios. The European Union's AI Act already includes security-related uses in the 'high-risk AI' category, so WMD information processing rules could be added during the implementation phase.

Deepening Specialization in AI Safety Talent Market
OpenAI's $455,000 salary setting signals the birth of a new professional category. A 'talent redeployment' phenomenon is expected to accelerate, with specialists from cybersecurity or threat intelligence fields moving to AI safety domains. Candidates with military service experience or defense research institute backgrounds are expected to be particularly favored.

Continued Balancing Act Between Transparency and Security
AI companies face a dilemma: they must externally demonstrate the effectiveness of safety measures while recognizing that disclosing specific defense mechanisms could provide attackers with circumvention pathways. This raises the need for a new governance model of 'verifiable opacity.' Solutions involving third-party auditing agencies or government certification systems serving as mediators are likely to be discussed.

Redefining Technology-Ethics Boundaries
The influx of WMD specialists into AI companies will serve as an opportunity to redraw the ethical boundaries of technology development. The focus of discussion is shifting from "what AI can do" to "what AI should know." This is a fundamental question that could ultimately influence the development direction of artificial general intelligence itself.