Daily Brief

The UK AI Security Institute benchmarked GPT-5.5 against Anthropic's Claude Mythos on offensive cybersecurity tasks, finding GPT-5.5 achieved 71.4% on expert-level challenges and performed comparably to Mythos overall. A multi-step attack simulation that required 12 hours of human expert time was completed by GPT-5.5 in 11 minutes at a cost of $1.73. This is a formal government evaluation, not a community benchmark.

Elon Musk confirmed in court proceedings that xAI "partly" used distillation from OpenAI models when training Grok. The admission surfaced during the ongoing Musk v. OpenAI litigation and directly contradicts xAI's positioning as an independent AI development effort. OpenAI's terms of service explicitly prohibit using outputs to train competing models.

A new paper comparing system-prompt self-orchestration against LangGraph, CrewAI, and the OpenAI Agents SDK on procedural tasks across 200 conversations finds that well-crafted prompts outperform external orchestration frameworks. The implication is that the complexity added by agent frameworks introduces failure modes without commensurate gains on structured procedural work.

The Qwen team released Qwen-Scope, a suite of sparse autoencoders (SAEs) covering the full Qwen 3.5 model family from 2B to 35B parameters. SAEs map internal model features to human-interpretable concepts, enabling targeted model editing, circuit analysis, and behavior steering without full fine-tuning.

IBM released Granite 4.1, an 8B open-weights model that benchmarks competitively against 32B-parameter models on code generation and reasoning tasks. The release continues the efficiency compression trend where each generation of small models absorbs the performance floor of the previous generation's large models.

Google DeepMind published research on an "AI co-clinician" model designed to augment clinical decision-making by integrating into existing healthcare workflows rather than replacing clinician judgment. The work focuses on AI-assisted diagnosis and treatment recommendation with human oversight structures built in.