AI news for Friday, August 21, 2026

Slack Code gives Claude Code, Devin, GitHub Copilot, and Vercel agents dedicated shared channels. A team can watch the plan, inspect diffs, see a live preview, and keep a human approval step before production—moving agent work out of one developer’s terminal and into a visible operating loop. Slack says Code works on any plan; access to the partner agent is still required.

In a worker-authored inquiry, Luciana and five colleagues at Spanish outsourcing firms serving Apple describe annotation quotas that could become mathematically impossible, active-time monitoring, language-based pay gaps, NDAs, and abrupt layoffs tied to Apple “headcount” decisions. These are their accounts, not independently audited findings. Their demand is direct: Apple should enforce labor protections through its AI supply chain.

404 Media documents images spreading on X in which a real photograph is only lightly changed—clothing, body shape, or another detail—rather than replaced with a visibly synthetic scene. Some carry Grok watermarks. Preserving most of the original photo can make the abuse look more credible than an obvious deepfake. X did not respond to the publication.

Publishers can now embed a Preferred Sources button that lets readers favor their coverage across Search, Discover, and Google News. Google says people are twice as likely to click a preferred source and have already selected 345,000 unique outlets. Those are Google’s figures, but the feature gives newsrooms a concrete—and unusually direct—audience lever to test now.

In VentureBeat’s survey, 85% of respondents used at least two agent-orchestration platforms and 64% used three; 21% had only reactive monitoring, with no real-time kill switch. The 107-enterprise sample is directional, not a census. Still, it exposes the operational question behind every agent rollout: can one policy actually stop work—and cost—across a fragmented stack?

Reuters reports that later this year Anthropic plans to let business customers retain data from covered models inside their own cloud environments. The company would still require 30 days of retention; the change is where the data sits, not how long it exists. It has not launched yet, but that distinction could remove a serious security objection for regulated deployments.

In July, Anthropic accounted for nearly 44% of Ramp customers paying either company, while OpenAI reached nearly 40% and grew faster early this quarter. The dataset spans more than 70,000 U.S. businesses, but it skews toward tech, counts customers rather than dollars, and is not the whole market. Nearly 56% of Ramp customers now pay for at least one AI product.

A peer-reviewed study finds that outputs from diffusion models are often unattributable to any individual training example—even when the model’s training set is known. Some generated images do reflect influential examples, but many arise from patterns distributed across the data. That complicates the intuitive story that every output has one findable original behind it, with consequences for provenance and copyright arguments.

Micro1 has reportedly climbed from a $100 million to a $500 million gross annualized run rate in eight months. But gross is the key word: the company passes roughly 60% to 70% of billings to experts doing the work, implying a net run rate closer to $150 million to $200 million. The source-based estimate is not booked revenue; it is a revealing map of where training-data money flows.

The plug-in works with iMessage, SMS, and RCS in ChatGPT Work and Codex on Apple-silicon Macs. Sending requires approval by default, but users can grant persistent permission per conversation; OpenAI’s guide flags revocation steps and a known issue involving tasks that disable approval prompts. It is useful today—and exactly the kind of agent permission worth checking twice.

Not directly: raw urine is too concentrated, corrosive, and unsanitary. But the joke points to a real solution. Municipal wastewater can be treated and reused in cooling systems instead of drinking water, and some facilities already do it. As AI pushes local water demand higher, the unglamorous infrastructure story is that sewers may matter almost as much as servers.