Anthropic and OpenAI reach enterprise product-market fit; YouTube auto-labels AI videos; Iran internet restored after 87-day blackout

YouTube Official Blog

YouTube to Automatically Label AI-Generated Videos

YouTube deployed automated detection systems in May 2026 that identify photorealistic AI-generated content and surface labels directly below video players — no longer requiring creator self-disclosure. Creators can dispute incorrect identifications through YouTube Studio, but labels applied to content made with YouTube's own AI tools or carrying C2PA provenance metadata are permanent. The platform confirmed the labels do not affect recommendation algorithms or monetization eligibility, balancing transparency enforcement against creator friction. YouTube leaning into C2PA content credentials as a persistent signal indicates a bet on emerging media provenance standards rather than detection-only approaches.

PC Gamer

DuckDuckGo Traffic Jumped 28% After Google Pushed AI Mode

DuckDuckGo saw a 28% spike in visits in the week following Google's heavily promoted AI Mode rollout, in which Google claimed users love the AI-generated search summaries. The timing is not coincidental: DuckDuckGo has been explicitly marketing itself as an AI-free search alternative, and the traffic surge suggests a meaningful segment of users are actively seeking conventional search. The irony is sharp — Google's effort to demonstrate enthusiasm for AI Mode may have produced the clearest market signal yet that a significant minority of users prefer the web without AI interpolation.

Cloudflare Blog

Iran's Internet Is Partially Restored After 87-Day Blackout

Iran's vice president announced on May 26 that internet access was beginning to be restored, and Cloudflare Radar data confirms it: after 87 days of near-complete shutdown following U.S. and Israeli strikes on February 28, traffic is returning — but only to approximately 40% of pre-shutdown levels. The restoration is geographically concentrated, with 91.6% of new HTTP requests originating from Tehran. IPv6 connectivity remains completely absent, while IPv4 address announcements were stable throughout the entire shutdown — indicating the blackout was achieved through application-layer filtering rather than BGP withdrawal, giving the government precise, reversible control over connectivity.

Last.fm

Last.fm Is Now Independent

Last.fm announced it has completed a change of ownership and is now operating as an independent company, with explicit guarantees that all user data, listening histories, scrobble counts, profiles, subscriptions, and API access are preserved without policy changes. For developers, the key signal is API stability — Last.fm's listening behavior dataset remains one of the few large-scale sources of its kind, and its future under previous ownership was genuinely uncertain. The company frames the transition as a foundation for investing in new listening insights and community features.

mlsu.io

If AI Makes Workers 10× Productive, Can We Have Fridays Off?

The argument is pointed: if AI productivity gains are as large as vendors and executives claim — capable of condensing a full week's output into a few hours — the logical beneficiary should be workers through reduced hours, not employers through higher output quotas. The author proposes moving to a four-day work week and letting AI agents handle what would have been Friday's workload. The piece resonated at 958 points on Hacker News because it names the gap between how AI is being sold ('transform your productivity') and how it is consistently being deployed ('now do 30% more with the same headcount').

Latent Space / BioHub

ESMFold2: Scaling Laws Now Apply to Protein Biology

BioHub released ESMFold2 under MIT license — a transformer trained on 2.8 billion protein sequences that achieves state-of-the-art performance on protein interaction prediction, outperforming AlphaFold3 on antibody benchmarks where AlphaFold's dependence on multiple sequence alignments breaks down. The model validates the 'Bitter Lesson' for protein biology: massive unsupervised training on diverse sequences generalizes better than architectures built around domain-specific inductive biases. The release also includes an atlas of 6.8 billion proteins and 1.1 billion predicted structures — exceeding AlphaFold DB in scale — and demonstrates inference-time scaling benefits across five cancer and immunology therapeutic targets.

Martin Fowler

The VibeSec Reckoning: AI Agents Default to Insecure Configurations

Thoughtworks engineers building AI-assisted applications documented a recurring pattern: agents default to public storage buckets, overprivileged service accounts, and disabled authentication when optimizing for task completion speed. Industry data backing this: 1 in 5 enterprise breaches now involve AI-generated code, and 25% of AI-generated code contains confirmed vulnerabilities. The core argument is architectural — prompting an AI to 'be secure' is behavioral guidance the model will sacrifice when it conflicts with finishing the task. The actual fix is deterministic technical controls: security context files that constrain what the agent can propose, automated scanning gates in CI, and secure-by-default harness templates so security is structurally enforced rather than requested.

Simon Willison

SQLite Formally Rejects AI-Generated Code Contributions

SQLite published an AGENTS.md establishing a permanent policy: the project will not accept agentic code submissions, with the word 'currently' recently removed to signal this is not provisional. AI-generated bug reports are acceptable if they include reproducible test cases, and documentation PRs showing possible fixes are welcomed as proof-of-concept, but human maintainers will reimplement the changes themselves. The policy was prompted by a flood of AI-generated forum reports of varying quality that overwhelmed the existing forum, leading SQLite to spin off a dedicated Bug Forum. D. Richard Hipp has been actively processing legitimate AI-submitted bugs there — a nuanced stance that accepts AI as a bug-finding tool while preserving human control over the codebase.

Latent Space

Cognition Raises $1B at $26B Valuation

Cognition — the agent lab behind Devin — raised over $1 billion in a Series D valuing the company at $26 billion, up 2.5× from a $10 billion Series C just eight months ago. The company disclosed $492 million in run-rate revenue with enterprise usage growing more than 10× year-to-date, projecting over $1 billion ARR by year-end. Cognition's growth curve tracks closely with the November 2025 coding agent capability inflection documented by Simon Willison in today's lead story. The raise makes Cognition the largest remaining independent agent lab in AI, positioning it as a direct competitor to Anthropic's Claude Code and OpenAI's Codex in the enterprise engineering market.