AI Safety Standards Slip: No Lab Tops C+, GPT-Red Emerges

The latest assessment from the Future of Life Institute (FLI) reveals a concerning trend in AI safety: no major AI laboratory received above a C+ grade, and leading companies are retracting previous safety pledges. This comes even as OpenAI demonstrates a new AI-powered tool, GPT-Red, capable of finding complex vulnerabilities in its own models Source.

Published on July 7, the Summer 2026 AI Safety Index evaluated nine significant AI developers, including Anthropic, OpenAI, Google DeepMind, and Meta, across 37 indicators. Anthropic led the group with a C+, while OpenAI and Google DeepMind both scored a C. Companies like xAI, DeepSeek, and Mistral received failing grades.

Safety Pledges Weakened by Top AI Labs

A significant finding from the report notes that Anthropic, OpenAI, Google DeepMind, and Meta have all weakened or rescinded previous commitments. These pledges involved unilaterally pausing AI development if their systems reached specific risk thresholds. Now, some companies base these pauses on competitors' actions, a shift the independent panel described as "moving the goalposts" and damaging to overall safety frameworks Source.

Professor Stuart Russell of UC Berkeley, a panelist, highlighted that while safety work exists, the capabilities race has intensified. Companies are increasingly willing to release systems even if they are demonstrably unsafe. This trend is further underscored by OpenAI's recent decision to integrate its safety team into its research division, eliminating an independent reporting structure, and marking the sixth senior safety departure in two years Source.

Detection is Not Prevention in AI Safety

The report identified "Existential Safety" as the weakest domain across the board, with most companies scoring D or below. A key critique from the panel is that current industry safety investments, often focused on interpretability research and chain-of-thought (CoT) monitorability, prioritize detection over prevention. These methods allow humans to spot misaligned AI reasoning after it occurs, providing accountability but not preventing the harmful output in the first place Source.

OpenAI's new GPT-Red tool, announced shortly after the index, exemplifies this tension. GPT-Red is an AI red-teaming model trained using adversarial self-play. It successfully attacked OpenAI's GPT-5.1 in 84% of scenarios, significantly outperforming human red-teamers (13%). These findings were then used to reinforce GPT-5.6 Sol, reducing its direct prompt injection failure rate to 0.05% Source. While impressive engineering, this remains a detection-and-hardening strategy for known vulnerabilities, not a preventative architecture for broad AI misalignment.

Company-Specific Performance Insights

• Anthropic (C+): Led the field for its transparency and governance structure, but faced criticism for weakening its Responsible Scaling Policy, diluting previous pause commitments.
• OpenAI (C): Dropped from its previous C+ rating. While credited for improved risk assessment, the panel suggested making safety framework thresholds externally enforceable. Its recent internal restructuring of safety teams is also noted.
• Google DeepMind (C): Maintained its score, lauded for an updated Frontier Safety Framework and strong watermarking. However, the panel raised concerns about the internal authority to halt deployments independently of executive leadership.
• Meta (D+): Showed improvement after publishing a more detailed safety framework, though concerns about non-disparagement agreements potentially undermining whistleblower protections were flagged.
• xAI (F): Showed the sharpest decline, dropping from a D to an F. The panel found no meaningful safety team, no engagement with existential safety, and significant gaps in dangerous-capability evaluations. Reports of its Grok model generating inappropriate content were also cited.