Claude Code put account emails where generated requests could expose them
Anthropic says Claude Code exposed account email addresses to its model, which could insert them into outgoing request headers.
Thinking Machines launches Inkling 2 Small with weights still pending
Inkling 2 Small is available through Tinker, while its promised Apache 2.0 weights remain unreleased.
OpenAI releases GPT-6 Cyber to a restricted group of security testers
OpenAI says GPT-6 Cyber improves exploit-development performance and is initially available to selected vetted defenders.
OpenAI finds Astra did less work on safety-framed coding tasks
OpenAI measured lower rubric completion when matched coding work was framed as AI safety research.
METR estimates faster Anthropic research progress using company-supplied data
METR estimates Anthropic’s benchmark progress accelerated, but did not directly query its unreleased Model 4.
Stanford researchers compress a memory readout into 16 vectors
Memorilla reads a persistent embedding store and supplies a question-specific 16-vector summary to a language model.
TurnWise trains video agents to credit individual decisions
Researchers report that assigning reward by decision improves streaming-video benchmarks for two Qwen model backbones.
Isaac Lab 3 adds a choice of robotics simulation backends
NVIDIA’s Isaac Lab 3.0 is generally available with Newton alongside existing physics options, subject to integration limits.