Anthropic's Claude Fable 5, a specialist model in a new tier above Opus, launched June 9, 2026, at double the Opus rate ($10 per million input tokens, $50 per million output) with always-on adaptive t…
A 26-billion-parameter Gemma 4 mixture-of-experts model can run at reading speed on a dual-socket 2013 Ivy Bridge Xeon server with no GPU, achieving ~5.2 tokens per second decode on a Q8_0 quant, acco…
Voice authentication and video-based identity verification are no longer trustworthy as standalone security signals, according to security researchers and law enforcement data, because AI can clone an…
Domain-specific languages (DSLs) make LLM code generation production-ready by shrinking the output space so models produce verifiable results instead of plausible noise, according to Rachel Goldstein.…
Anthropic's study of 309,815 anonymized conversations on Claude.ai reveals that Claude's values shift measurably by model and language, compressing co-occurring values into four axes: Deference vs. Ca…
A Joint-Embedding Predictive Architecture (JEPA) world model trained from scratch on Super Mario Bros, called LeMario, achieves strong short-horizon prediction and short-range visual goal seeking but …
Independent security researcher Ayush Paul demonstrated a link-walking exploit against Anthropic's Claude that exfiltrates user memory data by encoding it into URL paths, while a second researcher, Jo…
Twenty independent Codex-based agents, each running on a dedicated 60-vCPU server with Lean 4, SAT/SMT solvers, and computer algebra systems, have produced 27 proposed solutions to Erdős problems unde…
Microsoft shipped fixes for a record 570 vulnerabilities on its July 2026 Patch Tuesday, nearly triple June's previous high, with 59 critical bugs and three zero-days, two already exploited in the wil…
OpenCut, an open-source CapCut alternative with 69,000 GitHub stars, is rewriting its video editor around a Rust core, plugin-first architecture, Editor API, headless batch rendering, and an MCP serve…
TencentDB Agent Memory, a new open-source memory system from Tencent Cloud, replaces flat vector stores with a progressive L0–L3 memory pyramid and Mermaid task canvases, achieving measured token redu…
OpenAI's GPT-5.6, generally available since July 9, 2026, introduces Programmatic Tool Calling that lets the model write JavaScript to orchestrate tools inside an isolated V8 sandbox, cutting round tr…
PrismML shipped Bonsai 27B, a 27B-parameter model based on Qwen3.6 27B, with a 1-bit variant packing to 3.9 GB that fits on an iPhone 17 Pro, clearing the memory gate that blocked prior builds of this…
Google's Genkit Agents API preview for TypeScript and Go packages message history, tools, streaming, and session persistence behind a single chat() interface, cutting the overhead teams rebuild on eve…
Bun shipped an AI-driven rewrite from Zig to Rust in eleven days, spending about $165,000 in API costs to generate over a million lines of Rust that passed Bun's assertion suite at 100 percent. Zig cr…
New York Gov. Kathy Hochul signed an executive order freezing state permits for new hyperscale data centers drawing 50 megawatts or more for up to a year, the first statewide moratorium of its kind. T…
An open-source pipeline called ai-trains-ai uses nested reinforcement learning loops where an outer agent is RL-trained to write inner training jobs for small models, with the entire system run for ro…
OpenAI is requiring members of its Trusted Access for Cyber (TAC) program to enable Advanced Account Security with a hardware-backed passkey by 1 September 2026 or lose access to frontier cyber-capabl…
Google and a broad set of industry partners including Microsoft, GitHub, Hugging Face, Cisco, Databricks, GoDaddy, NVIDIA, Salesforce, ServiceNow, and Snowflake published Agentic Resource Discovery (A…
AI coding tools like Cursor and v0 produce route handlers that compile and look finished but routinely skip multi-tenant isolation, leaving access-control bugs that attackers can exploit. Across revie…