Anthropic paired a sharp reported jump in offensive cyber capability with a restricted distribution program, making access policy part of the model release itself. Elsewhere, AMD local inference broadened and a celebrity-amplified memory project exposed a recurring benchmark failure: marketing can detach a score from the task its code actually measures.
1. Anthropic routes Mythos through Project Glasswing
Anthropic introduced Claude Mythos Preview and limited access through Project Glasswing. In its own tests, Mythos produced ten full control-flow hijacks across roughly 7,000 OSS-Fuzz entry points and generated working exploits 181 times from Firefox vulnerabilities that Opus 4.6 exploited only twice in several hundred attempts.
The figures come from Anthropic's scaffold and mix model capability with search budget, orchestration, and a specialist validation team. Glasswing gives selected defenders an early window, but the actual security benefit sits downstream: candidate findings have to become reproducible reports and installed patches before wider access erodes that lead.
Sources: Anthropic's Mythos capability assessment · Anthropic's Project Glasswing · WIRED on the restricted program
2. Lemonade 10.1 widens AMD's local-model path
Lemonade 10.1 shipped for Windows and Linux with a simplified CLI, restored asynchronous model loading, API-key support, and Gemma 4 execution on AMD GPUs. The project also targets Ryzen AI NPUs and can fall back to x86-64 CPUs.
The release closes a compatibility gap more clearly than a performance gap. Support for an accelerator permits local execution, but memory bandwidth and model format set the usable speed. The open repository makes those limits measurable on the exact laptop or workstation where private inference is supposed to run.
Sources: Phoronix on Lemonade 10.1 · Lemonade repository
3. MemPalace's own notes undercut its perfect-score launch
MemPalace launched its AI memory system with claims of perfect LoCoMo and LongMemEval performance. Penfield's audit pointed to the repository's own notes: LoCoMo used a top-50 retrieval setting on conversations with at most 32 sessions, while the LongMemEval runner measured retrieval recall instead of end-to-end question answering.
Penfield develops a competing memory system, a commercial conflict that colors its framing. The methodological discrepancies remain directly inspectable in MemPalace's artifacts, however. The decisive comparison is metric identity: retrieval recall cannot be placed beside answer accuracy under one leaderboard label, even when both produce a percentage.
Sources: Penfield's competing-project audit · MemPalace repository and benchmark artifacts