Growing concern over the safety of rapidly advancing AI systems has intensified following the release of the Winter 2025 AI Safety Index, covered by Reuters, Red94 and NBC News. Together, the reports present a consistent picture: major AI companies are significantly behind emerging safety expectations at a time when model capabilities continue to accelerate.
According to Reuters, the Future of Life Institute (FLI) found that leading firms’ safety practices are “far short of emerging global standards,” and none of the assessed companies has a robust strategy for controlling advanced systems. Red94 similarly reports that none of the eight evaluated companies had “credible plans to control superintelligent systems”, with no firm scoring higher than a C+. NBC News reinforces this assessment, stating that companies “lack the concrete safeguards, independent oversight and credible long-term risk-management strategies that such powerful systems demand.”
A number of themes emerge across the sources. NBC News notes that experts identified two clusters of firms: Anthropic, OpenAI and Google DeepMind forming the leading tier, though still earning only a C+ at best, and xAI, Meta, Z.ai, DeepSeek and Alibaba Cloud forming a lower tier with more pronounced safety gaps. Red94 adds that experts found “zero credible strategies for maintaining human control” over highly capable systems, while also documenting present-day harms, including lawsuits referencing chatbot-linked suicides and reports of cyberattacks involving model exploitation. Reuters similarly highlights cases tying AI chatbots to incidents of self-harm, alongside concerns over lobbying against binding safety regulation.
Taken together, the reports illustrate an industry where capability expansion is occurring more quickly than the development of safety infrastructure. NBC News notes that capabilities continue to increase, with models like Google’s Gemini 3 and DeepSeek’s latest system performing at or near the frontier, while safety practices do not appear to be advancing at a comparable rate. Red94’s detailed breakdown of shortcomings across areas such as risk assessment, governance and information sharing further points to inconsistency in how companies approach safety. The repeated references to low grades, gaps in documentation and limited mitigation frameworks indicate that safety maturity varies widely across the sector.
From an industry-wide perspective, these findings suggest that organisations adopting AI systems may need to navigate uneven safety commitments across providers. This reflects what the reports describe: differing levels of transparency, varying governance structures and clear distinctions between the leading and lower-tier firms. The presence of documented real-world harms across multiple sources reinforces that safety considerations are increasingly relevant to operational and technology decisions.
As the reports collectively indicate, a key issue is whether governance and internal safeguards can evolve in line with system capability growth. The Safety Index recommends expanded transparency, reduced lobbying against safety measures and the greater use of independent evaluators, signalling areas of focus for organisations and regulators. Monitoring how companies respond to these recommendations, and whether the gap between capability and safety narrows, will be important in understanding how the sector progresses in the year ahead.