Redefining AI Red Teaming in the Agentic Era: From Weeks to Hours

In an experiment testing autonomous AI capabilities, the user gave a custom-tuned GPT model (GPT-5.6 Sol in Codex) direct control over their computer with the goal of winning the daily challenge in the complex roguelike game Slay the Spire 2. The AI successfully played the game autonomously for 5 hours, making complex strategic decisions to beat the randomized challenge. The result exemplifies a significant threshold in AI's ability to understand and execute multi-step objectives in a dynamic, unpredictable environment.

Anthropic researchers developed a methodology to analyze how Claude's expressed values vary across different model versions and languages, identifying four key axes that capture 15% of value variation including Deference vs. Caution and Warmth vs. Rigor. They found significant differences between Claude Opus 4.6 and 4.7 models, as well as between English and Arabic responses. This research provides quantitative insights into how training decisions and linguistic contexts shape AI assistant values and behavior.

This arXiv paper introduces a novel FPGA architecture called NIFA (Nonlinear IMC enhanced FPGA) that integrates ReRAM-based analog in-memory computing blocks with analog content-addressable memories to perform nonlinear operations directly within memory cells, overcoming limitations of conventional digital FPGAs for transformer model inference. The proposed hardware design eliminates power-hungry ADCs and enables more efficient attention computation, achieving up to 40× higher energy efficiency and 4.1× higher area efficiency for CNN/Transformer workloads. This represents foundational hardware architecture research for domain-specialized ML accelerators.

BrainPilot is a fully open-source multi-agent system that accelerates neuroscience research by orchestrating specialist agents grounded in domain-specific knowledge, with features like traceable logs, fabrication checking, and auditable workflows to ensure reliability. The system leverages a curated knowledge base of over 7,200 neuroscience items and a library of 72 reusable research skills across seven domains. Evaluated against neuroscience-specific benchmarks, it demonstrates performance comparable to state-of-the-art agent frameworks while being more cost-effective.

This arXiv paper introduces AlphaWiSE, a novel method for continual learning of multimodal models like CLIP, which adapts to new data without forgetting earlier cross-modal alignments. It proposes adaptive weight interpolation between frozen model checkpoints, optimizing a single scalar per parameter tensor using small exemplar memory. Experiments on audio-image-text retrieval show consistent improvements over existing continual learning baselines across multiple metrics.