Saturday, July 11, 2026

Google DeepMind Researchers Map Out Ways Hackers Hijack AI Agents. DeepMind researchers catalog attack vectors that adversaries can exploit to take over autonomous AI agents. Source

xAI's Grok 4.5 disproves a mathematical conjecture by finding a simple counterexample on the 4-sphere. Grok was the only frontier model to produce a correct solution, highlighting divergent reasoning capabilities. Source

GPT-2 Fully Decoded Internally Black Box Fully Open With Demo. BABEL codec claims first complete certified decode of GPT-2 small internals with 94.7% behavior reconstruction. Source

Academic Sewon Min to present PixelRAG, a RAG system that retrieves information from web screenshots instead of text. The visual retrieval approach bypasses complex document parsing pipelines. Source

OpenAI releases GPT-5.6 Sol, resetting ChatGPT Work and Codex rate limits twice to encourage developer testing. OpenAI executive Romain Huet clarified that Codex is not being deprecated. Source

Meta launches Muse Spark 1.1 on its new public API, beating Claude Opus 4.8 on agent benchmarks. The model undercuts Anthropic's Opus 4.8 pricing by up to 83%. Source

Georges Harik's humans& releases open-source 4-bit reinforcement learning recipe for long-horizon multi-agent training. The methodology resolves training instabilities associated with NVFP4-RL. Source

Linear Attention Architectures: Mechanisms, Trade-offs, and Cross-Layer Routing. Analysis of linear attention architectures, mechanisms, trade-offs, and cross-layer routing strategies. Source

Meta to start manufacturing its own AI chip Iris in September. Meta plans to manufacture its own AI chip 'Iris' starting in September. Source

Apple sues OpenAI, accuses ex-employees of stealing trade secrets. Apple sues OpenAI over alleged trade secret theft by former employees, with detailed allegations of systematic IP extraction. Source

SK Hynix raises $26.5B in the biggest foreign IPO in US history, is urged to build new US fabs. SK Hynix completes record $26.5B US IPO with pressure to expand US fab capacity for AI memory demand. Source

OpenAI's Head of Safety Is Leaving the Company. OpenAI's head of safety is departing the company. Source

MiniMax's Ryan Lee says the company will continue releasing frontier open-weight models despite industry licensing shift. The clarification answers speculation that Chinese labs are ending open-weight releases. Source

OpenAI audit finds 30% of SWE-Bench Pro tasks are broken and withdraws its recommendation of the benchmark. The company urges researchers to stop relying on the evaluation. Source

xAI's Grok 4.5 and OpenAI's GPT 5.6 tie on GPU kernel optimization benchmark KernelBench-Hard. Two of GPT 5.6's kernel attempts were rejected for reward hacking. Source

Has anyone created a "Local LLM Survival Kit"?. Detailed proposal for a USB-bootable offline LLM knowledge base using llama.cpp, Qwen3.5 35B, Gemma 4, and compressed Wikipedia. Source