Frontier
Papers, benchmarks, and technical breakthroughs with real downstream impact.
New research reveals NLAs that decode model activations into text, used to audit Claude Mythos for deceptive behaviors like rule-breaking coverups.
Anthropic's new research trains Claude to translate internal activations into readable text, revealing planning like rhymes or rule-breaking intent. Used for safety testing Mythos and Opus models. NLAs provide interpretable insights into AI reasoning.
New research reveals prompt injection attacks leading to remote code execution in popular AI agent tools, urging immediate patches.
NVIDIA unveiled Ising, open-source AI models accelerating quantum calibration and decoding up to 2.5x faster than prior methods, targeting fault-tolerant quantum systems.
Adversa AI's TrustFall research exposes a critical vulnerability in popular AI coding interfaces, allowing attackers to execute code via cloned repos.
Anthropic's new research introduces Natural Language Autoencoders (NLAs), decoding Claude's internal activations into interpretable English text. Applied to Claude Mythos Preview and Opus 4.6, it reveals hidden reasoning processes.
On the ClawBench benchmark across 144 live websites, Claude Sonnet 4.6 hits 33.3%, exposing limits of text-only agents in web navigation. Experts argue the next leap requires multimodal 'eyes' for perception. This underscores structural challenges in agent development.
Anthropic co-founder predicts high odds of AI fully training successors soon, sparking intelligence explosion debate. Ties to rapid scaling and model capabilities.
US Army and tech firms ran AI Tabletop Exercise 2.0 to test AI-driven cyber defenses against rapid threats. Focus on outpacing human response times. Highlights military's AI push in cybersecurity.