Daily Brief

The UK AI Security Institute benchmarked GPT-5.5 on offensive cybersecurity tasks and found it achieves 71.4% on expert-level exploitation challenges — performance the institute describes as the strongest they've evaluated to date. GPT-5.5 completed a multi-step cyber-attack simulation in 11 minutes at $1.73, a task that took a human expert 12 hours. Critically, GPT-5.5 slightly outperformed Anthropic's Mythos on at least one simulation, per community benchmarks.

OpenAI eliminated the provision in its Microsoft licensing agreement that would have voided the deal upon achieving AGI, substituting a fixed 2032 expiration date. The change enables multi-cloud licensing arrangements and sidesteps the governance ambiguity of defining AGI. It's a structural rewrite of the founding commercial relationship between the two companies.

In court proceedings related to the ongoing OpenAI litigation, Elon Musk confirmed that xAI "partly" used distillation from OpenAI models to train Grok. The admission surfaced during the trial and directly implicates xAI in the same training data sourcing questions it has leveled at others. The extent and methodology of the distillation remain unspecified.

The ARC-AGI-3 benchmark results for GPT-5.5 and Claude Opus 4.7 are out: 0.43% and 0.18% respectively. Both numbers are essentially noise-floor performance on a benchmark explicitly designed to resist pattern-matching and require novel reasoning. This comes the same week both models posted impressive results on cybersecurity and standard coding benchmarks.

PFlash, an open-source speculative prefill technique implemented in C++/CUDA, delivers a 10x prefill speed improvement over llama.cpp on 128K context windows with quantized 27B models running on a single RTX 3090. The implementation is open-source and targets consumer hardware constraints directly.

Graphify, a tool built as a Claude Code skill that uses Leiden community detection to generate knowledge graphs of codebases, reached 450K downloads and 40K GitHub stars in 26 days. The developer reports 71x token efficiency compared to raw file input. The tool was built solo in 26 days using Claude Code as the primary development environment.