AI Safety Audit Reveals Critical Gaps in Major Labs' Self-Checks

Security friction: AI labs show inconsistent self-safety practices

Guidelight, a nonprofit co-founded by former OpenAI safety leads, assessed internal AI safety practices at five major labs in August 2026. The analysis focused on six critical safety measures: logging activity, reviewing risky actions, enabling emergency shutdowns, and containing misaligned models. Results showed significant variation: Anthropic and OpenAI achieved a C+ grade, Google received a D+ with improvement plans, xAI scored a D-, and Meta got the lowest F grade. Crucially, no company fully implemented all required controls. The assessment relied solely on publicly available sources like system cards and reports, not internal audits or direct verification. This reveals current gaps in self-regulation that could enable security breaches or instability if unaddressed. The friction here directly threatens the security of systems that increasingly manage critical infrastructure and data. Next steps will depend on whether labs act on their published improvement roadmaps. The source’s public-only methodology means internal practices remain unverified—this assessment highlights risk but doesn’t confirm actual breach likelihood or scale. The material is brief and the detail sits with Guidelight’s report.

Source: The Decoder