All Articles

Proof-of-Control, a 127-requirement draft standard open for comment, grades agent evidence by who you must trust. The tiers, the six domains, what to do now.

A census of 20 AI agents by the identity each acts under: 11 get their own account or token, 9 reuse your login. What it means for revocation and blast radius.

IBM Research ran an agent five times per task: 77.4% of runs passed but only 53.0% of tasks passed every time. What the gap is and how to measure yours.

Since August 27, 2026 budget-limited Target CPA and Target ROAS campaigns bid to the target instead of beating it. What changed, which campaigns, five checks.

Anthropic's Life Sciences Verification Program gives vetted labs two grant types and moves enforcement from blocking to monitoring. Who qualifies, what it asks.

OpenAI's Astra for Law is GPT-6 Astra plus a legal search index, written instructions and gated access. What each part is, who gets it, what OpenAI measured.

List prices for one hour of audio or video across nine providers, from OpenAI and Google to Alibaba and the speech vendors, with every conversion shown.

28 on-record figures from 13 companies on how much of their own code or work AI does, each with its exact definition, date and source. Why none are comparable.

Tests check a few inputs; a proof covers all of them. Bend, Verus, Dafny and Lean compared on what you write, what they prove, and how an agent proves first.

A 4B model trained for $1,200 cut Postgres query latency 44.7% on a standard benchmark. The trait that made it work, and a routing table for your own tasks.

Apple's iPhone 18 Pro signs pixels at the sensor; C2PA Content Credentials record edit history. Who signs, what survives editing, and what to do now.

Researchers took over OpenAI staff ChatGPT accounts via a forum image bug and an SSO flaw, then reached internal repos via Codex. A checklist for connector use.

Anthropic says an AI agent made 30+ scientific models about 4x faster in four weeks, overseen by two staff new to kernel work. How to run it yourself.

Qwen's September 18 model takes audio and video natively with 1M context and, by Qwen's own method, cuts per-hour audio cost 98% versus its predecessor.

What an AI avatar video costs against a creator's, what platform and consumer data say about trust, which ad jobs suit each, and a 20-variant test worked out.

Every published rate of coding agents gaming their tests, with the count or population it is over: METR, ImpossibleBench and a September 2026 paper. 19 rows.

A September 2026 paper names four ways an agent workflow breaks a policy while every step passes its own check. The types, worked examples and the fix for each.

The 60 most-downloaded open text and vision-language models on Hugging Face, by 30-day downloads collected September 18, 2026. Qwen holds 27 rows; the caveats.

A UC Berkeley and Arena study ran seven models in three coding harnesses. Success barely moved; cost moved up to 5x. All 42 measured rows, with intervals.

Edge0 streams a 35B mixture-of-experts model from SSD in under 3 GiB of RAM. Bonsai 2 shrinks a 27B model to 5.9 GB. Which to pick, and what to test first.

Anthropic proposes three oversight metrics for AI agents and reports its own: 30,000 agents, 100% monitored, 1 in 47,000 blocked. How to measure yours.

GitHub will block the pull_request_target trigger in public repos from November 2, 2026. Who is affected, why AI review bots are exposed, and what to do.

A Stanford method published in Nature turns a paper and its code into tested tools an AI agent can call. The reported results, and the pattern for business.

SentinelLABS matched public Hugging Face commits to OpenAI's May agent incident, to the second. What the report shows, and the logs a platform should keep.