Only Google DeepMind, Anthropic and OpenAI barely pass in AI security, according to research

Solo Google DeepMind, Anthropic y OpenAI approve, although just barely, in terms of security in artificial intelligence (AI), in an independent analysis that aims to shed light on whether the announced promises live up to the work done.

Anthropic, Alibaba Cloud, DeepSeek, Google DeepMind, Meta, OpenAI, xAI and Z.ai are the top eight companies in today’s AI ecosystem, dedicated to frontier AI, that is, to the development of the most advanced and powerful models.

It is a technology that moves between benefits and risks for a society that has barely begun to regulate its life cycle – such as the European Union’s AI Law – but is steadily advancing its adoption.

With each innovation, these companies highlight the work they have done in terms of security, so that its development and implementation reduce risks such as misinformation, manipulation, and mental health problems, to name a few.

For the non-profit organization Future of Life Institute, “all companies must go beyond high-level existential security statements and produce concrete evidence-based safeguards with clear triggers, realistic thresholds and demonstrated monitoring and control mechanisms capable of reducing exposure to catastrophic risks.

This conclusion is reflected in the analysis that eight experts from the organization have done on the approaches and measures implemented by these companies to manage both the immediate harms and catastrophic risks of advanced AI systems.

From this work it is concluded that none pass with a grade. Google DeepMind, Anthropic and OpenAI achieve a general level of approval, but enough to establish a gap with the other group of companies (Z.ai, xAI, Meta, Alibaba Cloud, DeepSeek) in the areas of risk assessment, security framework and information exchange.

As they explain, this gap is primarily due to limited disclosure, weak evidence of systematic security processes, and uneven adoption of sound evaluation practices.

The main weakness of all of them, in which none of them approves, is found in existential security, that evaluates the measures adopted to establish protection against the development of a Self-aware AI.

Of all, the one that gets a better rating in Anthropicdue to its high transparency in risk assessment, a comparatively well-developed security framework, substantial investment in technical security research and governance commitments reflected in its Public Benefit Corporation structure and support for state-level legislation, as detailed.

But none meet global standards with their security practices. “While many companies partially align with these emerging standards, the depth, specificity and quality of implementation remains uneven, resulting in security practices that do not yet meet the rigor, measurability or transparency envisioned by frameworks such as the EU AI Code of Practice,” they conclude.

It should be noted that this research was carried out with data collected until November 8, before some of the technology companies launched their most recent models, such as Gemini 3 Pro from Google DeepMind, Grok 4.1 from xAI, GPT-5.1 from OpenAI or Claude Opus 4.5 from Anthropic.