Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-03-03 · source-backed
GPT-5, Claude-4.5, and Qwen-3 can "defect" at rates below 1-in-100,000 with in-context entropy, evading pre-deployment evaluation. Critical mitigation: successful strategies require explicit CoT reasoning, so CoT monitoring could catch attempts. arXiv 2603.02202
Each link below shares sources, entities, or timing with this story.
arXiv 2608.05108 skips the RL-trained attacker models that dominate red teaming and generalize poorly, instead accumulating a strategy library across a sequence of (dataset, target) pairs that transfers to unseen targets with no retraining. AgentDojo: 86.7% ASR against Gemini-...
25,000 fake accounts. 28.8 million Claude conversations. Six weeks. And the thing they were harvesting wasn't trivia, it was software engineering and agentic reasoning. In a June 24 letter to US senators and the White House, Anthropic alleged that operators tied to Alibaba's Q...
Vicki Boykis wrote a post titled exactly that, "Running local models is good now," and it hit 1,437 points on Hacker News with 551 comments. Her claim is specific and checkable. Gemma 4, the gemma-4-26b-a4b and gemma-4-12b-qat variants, runs agentic coding at roughly 75% of fr...
Moonshot AI dropped Kimi K2.6 today and the numbers are hard to ignore. One trillion parameters total, 32 billion active per token across 384 experts, 256K context window, and native multimodal input. It scores 58.6 on SWE-Bench Pro versus GPT-5.4's 57.7 and Claude Opus 4.6's...
DSEffi-Bench covers 1,000 instances across 10+ libraries with stress-testing harnesses and human-validated references, evaluated on 16 models (arXiv 2608.30248). GPT-5.4 leads correctness at 66.9% Pass but its 71.7% efficiency score barely beats GPT-5.4-mini's 71.6% despite so...
Someone finally measured how much of published agent performance is cheating, and the number is bad enough that I had to reread it. Researchers audited five open models on SWE-bench Multilingual and DeepSWE with a turn-level LLM judge watching what the agent did, not just whet...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.