Risque de développement d'armes biologiques et chimiques facilité par l'IA

Frontier AI systems are demonstrating capabilities relevant to biological and chemical weapon development that multiple AI developers have been unable to confidently rule out as providing meaningful uplift to non-expert actors. In May 2025, Anthropic activated ASL-3 protections — its second-highest safety tier — for Claude Opus 4 because it could not "confidently rule out the ability of their most advanced model to uplift people with basic STEM backgrounds" for bio/chem weapons development. This was the first time any AI developer deployed a model under its highest activated safety level specifically due to biosecurity concerns.

In June 2025, RAND Corporation tested three frontier models (Llama 3.1 405B, ChatGPT-4o, and Claude 3.5 Sonnet) and found that all three "successfully provide accurate instructions and guidance for recovering a live poliovirus from a construct built from commercially obtained synthetic DNA." RAND argued that existing safety assessments "underestimate this risk" due to flawed assumptions about the tacit knowledge barriers that remain. The IASR 2026, chaired by Canadian researcher Yoshua Bengio, reports that "in one study a recent model outperformed 94% of domain experts at troubleshooting virology laboratory protocols" — referring to OpenAI's o3 achieving 43.8% accuracy on SecureBio's Virology Capabilities Test versus 22.1% average for human experts in their sub-specialties.

A separate line of evidence concerns AI protein design tools. In October 2025, research published in Science found that AI protein design tools generated over 70,000 DNA sequences for variant forms of controlled toxic proteins, and one screening tool missed more than 75% of potential toxins. After a 10-month remediation effort, screening was improved to catch 97% of high-risk sequences — but the episode demonstrated that biosecurity controls have not kept pace with capability.

Canada's exposure to this risk is direct and multi-layered. Canada hosts the sole BSL-4 facility in the country — the Canadian Science Centre for Human and Animal Health in Winnipeg — which works with Ebola, Marburg, and other high-consequence pathogens. The Qiu/Cheng incident (scientists terminated 2019-2021, CSIS investigation confirmed 2024) demonstrated that insider threats at this facility are real: CSIS found intentional transfer of scientific knowledge and materials related to Ebola and Marburg viruses. Canada has 17+ academic BSL-3 facilities through the CCABL3 consortium. VIDO at the University of Saskatchewan is constructing a second BSL-4, which will be the only non-government CL4 in Canada.

Canada's Sensitive Technology List (published February 2025) explicitly identifies the convergence: "advancements in nanotechnology, synthetic biology, artificial intelligence and sensing technologies could provide enhancements to existing weapons, such as biological/chemical weapons." Canada signed the Seoul Declaration (May 2024) recognizing that frontier AI could "meaningfully assist non-state actors in advancing the development, production, acquisition or use of chemical or biological weapons."

Yet Canada has published no dedicated assessment of AI-enabled bio/chem weapon risk. The Canadian AI Safety Institute (CAISI, launched November 2024, $50M over five years) does not yet explicitly include biosecurity evaluations in its public mandate. The gap between Canada's international commitments acknowledging this risk and its domestic institutional capacity to evaluate it is the core governance concern.

Anthropic's activation of ASL-3 protections — the first such deployment by any AI developer — represents a case where voluntary safety frameworks functioned as designed: the company identified a risk during pre-deployment evaluation and applied its highest activated safety tier in response. Several other frontier AI developers have also implemented pre-deployment biosecurity evaluations. The debate centers on whether voluntary measures are sufficient or whether mandatory requirements are needed to ensure consistent evaluation across all developers.