Not a single AI company passes safety test: FLI report reveals systematic retreat from safety commitments
What it really says
The Future of Life Institute (FLI) released its semi-annual AI Safety Index on July 7, 2026. An independent panel of seven experts from institutions including the University of Montreal, UC Berkeley, and Oxford evaluated nine leading AI companies across six categories: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and transparency and communication. The results are sobering: not a single company received an A or B grade. Anthropic scored highest with C+, followed by OpenAI and Google DeepMind at C each. Meta received only a D+, Z.ai and Alibaba Cloud each received D-. The worst grades went to xAI, DeepSeek, and French company Mistral - all three received an F. Most concerning: the experts found that companies have systematically weakened or abandoned their own voluntary safety commitments. So-called 'red lines' - pledges to halt development when AI systems approach dangerous capability thresholds - have been internally diluted. Policies prohibiting military use of AI have also been successively withdrawn. FLI Chair Max Tegmark commented: 'AI companies are sprinting toward a cliff.'
Our assessment
This story merits a red rating because it documents a concrete and verified governance failure. When the world's leading AI companies systematically weaken their own voluntarily adopted safety commitments - while their models simultaneously grow more capable - that is a serious warning signal. The fact that not a single company achieved a good grade shows the breadth of the problem. The finding about abandoned 'red lines' is particularly significant: these pledges were a central industry argument against stricter government regulation. Their erosion strengthens the EU's position of mandating binding rules through the AI Act rather than relying on voluntary self-regulation. However, methodology should be considered: the FLI is an organization that advocates for strict AI regulation, and evaluation criteria - particularly for 'existential safety' - go beyond what even the EU AI Act requires. The low grades thus partly reflect a very high standard. Nevertheless, the trend away from safety commitments is real and confirmed by multiple independent reports.
Relevance for Germany
This report is highly relevant for Germany and Europe for several reasons. First, Mistral, the only European AI company evaluated, received the worst grade of F - a finding that burdens the debate about European AI sovereignty, since safety should be a European competitive advantage. Second, the report appears less than four weeks before August 2, 2026, when the EU AI Act's transparency obligations take effect, underscoring the need for binding regulation. Third, the Bundesnetzagentur as Germany's AI supervisory authority and the BSI, which just published its A5 audit catalog for trustworthy AI, can use the FLI findings as evidence that voluntary self-regulation is insufficient. Fourth, for German companies deploying AI, the question arises whether their providers maintain reliable safety standards - the FLI index provides initial orientation.
Fact check
The FLI Safety Index from July 7, 2026 is confirmed by numerous independent sources. Axios, Time, Seeking Alpha, Seoul Economic Daily, Malay Mail, and TechXplore report consistently on the findings. The grades (Anthropic C+, OpenAI C, Google DeepMind C, Meta D+, Z.ai and Alibaba Cloud D-, xAI/DeepSeek/Mistral F) are consistently cited across all sources. The seven-member expert panel from the University of Montreal, UC Berkeley, and Oxford is named in the reports. Max Tegmark's quote comes from the Axios report. The finding about retreating from safety commitments and lifting military use prohibitions is independently confirmed by Axios and Seoul Economic Daily. The six evaluation categories are documented in multiple sources.
Source
- • https://www.axios.com/2026/07/07/report-ai-safety-pledges
- • https://time.com/article/2026/07/07/ai-safety-rankings-openai-anthropic-meta/
- • https://seekingalpha.com/news/4611964-google-anthropic-openai-lead-ai-safety-index-spacexai-receives-f
- • https://en.sedaily.com/technology/2026/07/08/global-big-tech-retreats-on-ai-safety-pledges-experts-warn