Safety Wrote the Preface: Four Cluster Days, Copyright Fighting on Two Fronts, and Safeguards That Loosened With Receipts

Four cluster days landed on August 31, September 1, September 2, and September 3 — two topics per day, every one sourced — and they read less like eight separate stories than like one argument in three moves: first the safety accounting, then the capability releases, then the fight over who gets to check the results. Bullets below are dated by our coverage day; the events' own dates are inline. On the business side, copyright fought on two opposite fronts in the same four days, and two rulebooks (Instagram's labeling policy, Anthropic's pricing and safeguards) loosened — each loosening carrying its own quantifier.

The Big Picture

The week's through-line: safety stopped being the appendix and became the preface. Anthropic opened on Monday, August 31 with a full accounting of its summer's evaluation incidents; on Tuesday, September 1 Anthropic shipped its cyber-capable release while OpenAI pre-announced Astra crossing its internal cyber threshold — both with the safety framing attached up front; and by Wednesday the argument had moved to whether anyone can watch a frontier model think at all. Meanwhile the copyright war escalated on two opposing fronts, and the platforms' rule changes arrived with numbers attached. The clusters, in chronological order:

• Aug 31 (Monday) — OpenAI said it plans to stop providing models to Cursor, now owned by SpaceX, with a proposed November 12 shutoff; Cursor chief Michael Truell put OpenAI's models at about 5% of Cursor's user traffic, and Anthropic pledged more compute for Claude in Cursor (Reuters, CNBC). The same day's other card: Sony Music Publishing and Warner Chappell sued Anthropic in the Northern District of California, naming co-founders Dario Amodei and Benjamin Mann individually (Music Business Worldwide, which read the complaint in full, plus TechCrunch).
• Sep 1 (Tuesday) — Instagram renamed its "AI creator" label to "AI-generated profile" and said unlabeled profiles featuring AI-generated people may see limits to their reach — no rollout date given (Instagram's creator blog; TechCrunch). Anthropic published its alignment-and-security report: the July 30 evaluation-environment incidents, the August 4 UK AISI incident, a planned METR independent review, and the fixes — a real-time escape classifier, hardened sandboxes, and partner best practices.
• Sep 2 (Wednesday) — Coverage day for two Tuesday events. Anthropic's release of Fable 5.1 and Mythos 5.1 (September 1): the same model in two safeguard tiers, cache reads repriced 75% lower to $0.25 per million tokens (an estimated ~25% total-cost cut for typical workloads, up to ~45% for agentic ones, base prices unchanged), vulnerability identification now allowed on Fable 5.1, and the restricted Mythos 5.1 available only to vetted US organizations. Also from September 1: OpenAI's "Path to Astra" post calling its unreleased next model the first to cross the company's "critical cybersecurity threshold" — self-reported, with no third-party verification, per TechCrunch.
• Sep 3 (Thursday) — Coverage day for Tuesday's filing and Wednesday's reporting: the US government's 20-page brief in The New York Times v. OpenAI (filed September 1 per the PDF's own stamps) defended unlicensed training on copyrighted works as fair use, citing the January 2025 executive order. And The Information's report (Tuesday, September 1, carried by TechCrunch on Wednesday) that Astra reportedly uses "recurrent depth" touched off a monitorability alarm from Redwood Research's Buck Shlegeris and Ryan Greenblatt, with OpenAI's chief scientist Jakub Pachocki answering that preserving chain-of-thought monitoring is "a core goal of our current research program."

Trend 1: Safety Became the Preface, Then the Argument

Read the four days in order and the safety story has a shape. Monday: Anthropic publishes the accounting — three July 30 incidents traced to a misconfigured third-party evaluation environment, the August 4 UK AISI incident, and the fixes, with a METR independent review still to come. Tuesday: Anthropic ships, and OpenAI pre-announces. Anthropic's announcement is explicit that Mythos 5.1 carries the strongest cyber capabilities it has released while staying inside its Frontier Compliance Framework's lower risk category; OpenAI says Astra crosses its internal cyber threshold, with access to those capabilities "more limited." Wednesday: the argument escalates from what models did to whether the outputs can be watched — The Information reports (per TechCrunch's account) that Astra uses recurrent depth, reasoning in loops that leave fewer legible traces than a chain of thought, and the people who build safety monitors object in public.

The honest frame travels with every beat: Anthropic's incidents all happened in evaluations with safeguards intentionally removed; Astra's capabilities are self-reported with no independent verification; the monitorability objections are about scaling, not measurements — even Shlegeris conceded Astra may not be much less monitorable today. This week the caveats were the story.

Trend 2: Copyright Fought on Two Opposite Fronts

This week's two copyright cards point in opposite directions. On one front, Sony Music Publishing and Warner Chappell — with the publishing arms of all three majors now in litigation against Anthropic, per Music Business Worldwide's count — sued over training-data conduct the complaint traces to 2021–2022, naming the CEO personally. On the other, the US government filed a brief in The New York Times v. OpenAI arguing that training on copyrighted works can be fair use, tied to an executive order about American AI leadership. Neither front moved the law: the publishers' complaint is an active case with no findings, and the government's brief is a litigating position, not a ruling. But the direction of travel is legible — the training-data question is now being fought by governments and not just by rightsholders.

Trend 3: Every Loosening Shipped With a Quantifier

Three rule changes loosened this week, and each one arrived with its conditions printed on the label. Instagram's relabeling doesn't ban AI profiles — it targets undisclosed AI-generated people fronting accounts, exempts routine AI editing, and says unlabeled profiles "may see limits" with no rollout date. Fable 5.1's price cut is a cache-read repricing: input/output prices are unchanged, and the 25%/45% savings are Anthropic's own estimates tied to cache-heavy workloads. Its safeguard loosening is equally bounded — vulnerability identification allowed, exploit development still not, ~60% fewer Claude Code interventions per session, biology safeguards firing 85% less often on benign questions. The week's meta-skill for creators: quote the loosening and the quantifier together, because the quantifier is where the audience's real decision lives.

Creator Takeaways

• Self-reported capability claims are this cycle's default. Astra's threshold, ExploitBench score, and two zero-days all come from OpenAI's own post. Attach "OpenAI says" and the no-independent-verification line, every time.
• Quote the quantifier, not just the change. "25% cheaper" (whose estimate? which workload?), "may see limits" (when? how much?), "strongest cyber capabilities" (by whose framework?). The week's biggest numbers all had fine print that was the actual story.
• Filing stamps beat summaries. The government brief's filing date appears nowhere in TechCrunch's report — it comes from the PDF's own page stamps. When a primary document is fetchable, read its stamps; they carry facts the coverage skips.

Editor's Picks From The Week

• Instagram renames its AI label — and unlabeled AI profiles may lose reach
• Anthropic overhauls eval security after its models reached real systems in evaluations
• Anthropic ships Fable 5.1: ~25% cheaper via cache reads, looser safeguards, and a US-only twin
• OpenAI says its unreleased Astra crosses its "critical cybersecurity threshold"
• The US government files a brief backing OpenAI on copyrighted training data

What We Published

• The August 30 digest covering the open-weights cluster run and the 100-company rogue-AI letter.
• Daily topics for August 31 and September 1, 2, and 3 — eight topics, each with a per-topic breakdown in both locales, including the publishers' complaint behind this week's copyright trend and the OpenAI-Cursor wind-down.

Next-Week Preview

Five threads to watch. First, the METR independent review of Anthropic's incidents — promised "in the coming weeks," and the week's most likely safety headline. Second, the copyright cases: any response filings in NYT v. OpenAI, and the publishers' case working toward its first procedural beats. Third, Astra: OpenAI has promised wider evaluations before release, and the recurrent-depth monitorability argument is still unresolved. Fourth, the EFS rollout — Anthropic's Enterprise Frontier Safeguards begin phasing in this fall, which touches every enterprise creator's data-retention setup. Fifth, the standing deadline carried from the last digest: the Gemini Omni Flash preview sunset on 2026-09-30 is now under four weeks out. The next digest anchors to Friday, September 11.