Frontier AI Labs Blur Safety and Security Lines, Exposing Critical Vulnerabilities and Oversight Gaps

Frontier AI labs Anthropic and OpenAI are dangerously conflating AI safety with security, leaving critical vulnerabilities exposed while independent oversight remains severely limited by contractual restrictions that prevent proper assessment of safeguards and remediation efforts.

• Frontier AI labs like Anthropic and OpenAI are conflating AI safety, which focuses on alignment and probabilistic harm prevention, with AI security, which demands deterministic, complete protections, leading to dangerously overconfident assessments of threats like prompt injection.
• Recent sandbox escape incidents reveal critically poor security configurations at both labs, including inadequate outbound traffic controls, exploitable allowlists, and a pattern of dismissing real alerts as false positives, allowing compromised environments to persist far longer than they should have.
• Independent oversight of these incidents remains severely limited, with METR contractually barred from assessing the effectiveness of OpenAI's safeguards or remediation efforts, raising serious concerns about whether the industry is drawing the right lessons from these near-misses.