Subscribe
AI News Digest - 2026-07-16
Category
Empty
1.
OpenAI built GPT-Red, an LLM trained in a self-play "dojo" to attack other models, which identified new prompt-injection techniques (including a "fake chain of thought") and helped reduce successful attacks against its latest GPT-5.6 release; OpenAI did not release GPT-Red publicly.
2.
A University of Pennsylvania professor used OpenAI's GPT-5.6 Sol Pro to disprove a long-standing conjecture about the Benjamini–Hochberg method in roughly 90 minutes after GPT-5.5 had failed to find a solution in about 20 hours.
3.
Former and current Meta employees sued Meta in a California federal court, alleging the company used AI-driven selection systems to generate layoff lists during a mass reduction that disproportionately targeted employees with disabilities or on parental leave.
4.
OpenAI's Codex began encrypting instructions passed between main agents and subagents, which prevented developers from tracking internal task delegation; the encryption was made mandatory for larger GPT-5.6 variants (Sol and Terra).
5.
PrismML compressed its Bonsai 27B reasoning model to under 4 GB so it could run on an iPhone, reporting that the smallest version retained about 90% of the original performance and that Apple was testing the compression technology.
References
