Experts voice concerns about AI use in developing biological weapons

Artificial intelligence (AI) systems possess expert-level knowledge of developing biological and chemical weapons, but threat actors face significant physical barriers, according to the International AI Safety Report.

The International AI Safety Report has addressed the risks of using AI in biological and chemical weapons development, as AI systems now match or exceed expert knowledge relevant to their creation.

However, material barriers, such as acquiring equipment and obtaining regulated materials, remain significant challenges for novices seeking to develop weapons.

Accurately assessing AI capabilities in assisting the development of chemical and biological weapons remains challenging due to legal constraints and international treaties.

For instance, researchers who carry out or publish results on studies examining AI assistance in weapons development may inadvertently violate national security laws or treaties such as the Biological Weapons Convention.

As a result, researchers often rely on “benign proxy tasks.” They test how much AI systems may help with similar but harmless procedures, such as culturing low-risk bacteria.

While information related to the creation of chemical and biological weapons is already accessible online, AI systems allow novices to “access and contextualize relevant information faster than they could with internet searches alone.”

In 2025, several developers attempted to mitigate risks by implementing safety controls that reject harmful requests.

Safeguards include excluding pathogen data from training and restricting access to high-risk tools, but they haven’t been thoroughly tested and can be removed from open-weight models.

The measures to fight deepfakes are limited

The report also states that accessible AI tools have substantially lowered the barrier to creating harmful synthetic content at scale, which can be used for malicious purposes like fraud, extortion, and harassment.

Deepfake pornography is a particular concern, as 96% of deepfake videos online are pornographic and disproportionately target women. However, due to a lack of systemic data, the overall risk is hard to assess, and effective interventions are difficult to design.

Measures like AI warning labels yielded mixed results. In one study, they improved participants’ accuracy at identifying AI-generated videos from 10.7% to only 21.6%.

Prevention-focused techniques, such as limiting access to AI models to vetted users or imposing safeguards to prevent models from generating harmful content, can be bypassed.

Meanwhile, filtering sexual content from models’ training data is emerging as an effective method for increasing barriers to generating non-consensual intimate imagery.

The report states that watermarking and content logs are promising methods for verifying content authenticity; however, skilled actors can remove standalone watermarks or deceive detectors, reducing their effectiveness.

Curious what others think about this story? Contribute your thoughts to the debate below.

Therefore, the authors propose embedding watermarks or securing metadata in authentic media. For instance, recording devices can be required to embed unique digital signatures that help distinguish recordings made using them from synthetic content.

Another approach involves maintaining logs of AI outputs and using them to identify newly created synthetic content by comparison.

“However, this approach faces scalability issues, is vulnerable to evasion, and raises privacy concerns related to logging user interactions,” the authors write.

AI under scrutiny

The report comes amid ongoing investigations in France and the UK against X, the US-based social media platform, where users have been sharing Grok-generated nonconsensual images of women and minors.

The annual report, first released in 2025, was initiated by Professor Yoshua Bengio, often referred to as “the Godfather of AI” and the most-cited living scientist in the world.

The writing team consisted of over 100 independent experts from more than 30 countries and intragovernmental organizations, including the European Union (EU), the Organization for Economic Co-operation and Development (OECD), and the United Nations (UN).

The analysis explores the dangers of “general-purpose AI” – models and systems that can perform a wide variety of tasks and are used by nearly a billion people in their daily lives.

Unlock more exclusive Cybernews content on YouTube