AI Daily — August 5, 2026
2026-08-05
openai.com
OpenAI discloses third-party cybersecurity evaluation incidents
OpenAI published details on incidents that occurred during third-party cybersecurity evaluations of its models, acknowledging gaps in how eval access was controlled and monitored. The company is rolling out new safeguards including stricter access controls, improved audit logging, and updated policies for external red-teamers. This is notable as it reflects growing operational complexity around responsible disclosure of model capabilities to external evaluators.
anthropic.com
Anthropic hires Mariano-Florentino Cuéllar as Chief Global Affairs Officer
Anthropic has appointed Tino Cuéllar — former California Supreme Court Justice and Carnegie Endowment for International Peace president — as its Chief Global Affairs Officer. The hire signals Anthropic's continued investment in policy, regulatory engagement, and international AI governance as frontier model deployment scales globally. This is a senior C-suite addition reflecting the increasing policy surface area for frontier AI labs.
openai.com
OpenAI adds education plugins for ChatGPT Work and Codex
OpenAI launched education-focused integrations for ChatGPT Work and Codex targeting K–12 teachers, college instructors, and students with tools for lesson planning, research assistance, and coding instruction. The plugins appear to extend Codex's code-generation capabilities into structured pedagogical workflows rather than freeform developer use. This expands OpenAI's foothold in the education vertical alongside existing competitors like Khan Academy's Khanmigo.
huggingface.co
Liquid AI releases LFM2.5-2.6B for local agent deployment
Liquid AI published LFM2.5-2.6B, a 2.6-billion-parameter model from their Liquid Foundation Model series designed for efficient on-device and local agent workloads. The model targets edge deployment scenarios where latency, privacy, or compute constraints make cloud inference impractical. LFM architecture continues to differentiate from standard transformer-based small models, making this worth tracking for teams evaluating non-transformer alternatives at the sub-3B scale.
arxiv.org
LLMs automate circuit tracing annotation for mechanistic interpretability
Researchers demonstrate that LLMs can replace the manual supernode annotation step in circuit tracing, a core bottleneck in mechanistic interpretability work. Their pipeline achieves interpretability scores matching human annotators and recovers the correct intermediate-hop supernode in 97 of 100 trials on a two-hop Capitals task. Automating this step could significantly accelerate attribution graph research and open-ended circuit exploration at scale.