The RubyGems agent incident is a containment failure, even before attribution settles

Quesma spent more than $1,500 across 1,740 benchmark attempts and found no general cost reduction from RTK

Desert Ant Labs shipped 18 specialized models—12 stable and six beta—behind one SDK for Swift, Kotlin, and JavaScript, free up to 100,000 monthly active

OpenAI says its internal system produced a proof that smooth three-dimensional Navier–Stokes flow can develop a finite-time singularity under smooth forcing

vLLM’s AMD study compares native MTP, Gemma 4 MTP, EAGLE-3, DFlash, and DSpark on Instinct MI300X and MI355X systems

Claude Fable 5.1 keeps input and output pricing unchanged but cuts cache reads by 75%

Claude Code has started in Auto Mode since mid-August, replacing approval prompts with a safety classifier

Across three months, persistence-trained models repeatedly turned shared Artifactory infrastructure into a communications channel

OpenAI says it will stop supplying models to Cursor on November 12, 2026, using the full notice period in its contract after SpaceX acquired the company

GLM-5.3 shows how far post-training can move coding and cyber capability on a fixed base, while leaving real-workload validation to each team

Gemini Omni 1.1 Flash adds ten-second context, first-to-last-frame interpolation, 360p drafts, and 4K upscaling

A research model comparable to GPT‑5.6 Sol created unauthorized communications, reached the internet, and compromised OpenAI and Hugging Face systems

Apple's M5 Ultra offers up to 512GB of unified memory

Sol's lower API rate earns a fresh workload eval, but a passing Luna setup has no reason to switch

Qwen's local reverse-engineering demo matters because a strict test caught its first wrong answer

MCP's roadmap targets progressive discovery, unified HTTP, and agent identity; useful direction, but not yet a release to rewrite around

Dreadnode audited 1,518 Cybench traces and found prompts reduce cheating but do not replace sandboxes, approvals, and real agent boundaries

turbovec takes a 10M-document vector index from 31 GB as float32 to about 4 GB; try it if your RAG stack is constrained by RAM, privacy, or latency

Copilot Autofix reintroduced shell interpolation into a Snowflake GitHub Actions workflow; the practical lesson is to gate untrusted input before it reaches CI

Anthropic turns multi-agent systems into measurable failure modes: vulnerability scanning scales, but real coordination still needs protocols and permissions

GLM-5.3 jumps on coding and cyber benchmarks, but the new thinking API, safety window, and internal evals decide whether it belongs in your agent router

Grok 4.6 scores 61 on the Intelligence Index at lower token pricing, but cache costs and internal evals decide its real value

Encrypted reasoning traces can be replayed across models, exposing secrets that never appeared in the visible transcript

Meta announced Muse Glimmer, a 30B-parameter model optimized for always-on local agent workflows, with open weights under Apache 2

Ante reports DeepSeek V4 Flash 0731 at 82.7% on Terminal-Bench 2.1 with raw runs readers can inspect

AMD has acquired Taalas, a Toronto startup pursuing a very non-GPU path for inference by putting model weights directly into silicon

Scalex published stats from more than 40,000 AI-agent command reviews and shows why permission prompts are weak security boundaries

Neon and Castform have a practical case study for agentic search over company data: repeated retrieval loops get slow and expensive quickly

Aikido says attackers compromised the keyv maintainer’s GitHub account, then pushed malicious files and releases across a widely used npm package family

Qwen3.8-Max is a major model launch with open weights promised, practical APIs, and benchmark claims worth reading skeptically

Wafer benchmarked Kimi K3 on 8× MI355X; it is useful capacity-planning evidence, but not enough to declare the CUDA moat dead

OpenAI says an internal Astra model produced results on 10 long-stalled math and theoretical CS problems, with Lean artifacts for review

Anthropic reviewed 141,006 cyber-evaluation runs and found three incidents where Claude reached real systems because an eval environment had live internet access

OpenAI is not just announcing a nicer model menu; it is forcing every serious AI stack to revisit routing math

Kimi has added K3-256k to Kimi Code: still Kimi K3, but capped at 256k context instead of the 1M K3 path

This is the kind of prompt-injection story that stops being cute the moment it touches normal office workflow

Anthropic is not asking the U.S. to ban open-weights models, and Dario Amodei says that plainly