Archive — 2026-08Week 5

Bloomberg reports Anthropic's mega-IPO plan looms over a packed US listing calendar — by Bailey Lipschultz and Anthony Hughes, republished by Yahoo Finance, published 2026-08-31 13:02 UTC. Anthropic is expected to file for a public offering within weeks, raising as much as SpaceX's record $86.2 billion June debut or more, and to list ahead of OpenAI, which now eyes 2027. Companies in the queue are struggling to win investor attention in the post-Labor Day (September 7) window, with some investors and sovereign wealth funds delaying engagement with other IPO candidates. Named companies adjusting timing: cloud provider Nscale, smart-ring maker Oura Health, temporary-power provider Aggreko, solar-and-storage firm CoVolt Power, SB Energy targeting over $5 billion, Inspire Brands weighing a pre-election listing, and Switch, confidentially filed at a potential $50 billion valuation. Barclays' Americas ECM head Rob Stowe says activity will concentrate into a few windows (mid-September Fed meeting and the early-November election as hurdles); Morgan Stanley's Eddie Molloy notes non-AI IPOs face headwinds while "AI is the flavor of the month." Precedent: the 14 companies that listed before SpaceX's debut posted a weighted average loss of 9.5%.

Business Today relays The Information reporter Aaron Tilley's 2026-08-30 report: OpenAI has bought tens of thousands of Mac mini and Mac Studio systems for reinforcement learning (RL) and for training AI agents that autonomously operate computers, while Anthropic rents Mac minis through Amazon Web Services (AWS) for similar workloads — Business Today filed on 2026-08-31 at 12:08 PM IST (06:38 UTC). Both companies use the Macs for computer-use agent training in which the agent runs inside a real operating system, observes the screen, takes actions and receives feedback millions of times over; this is separate from the leased cloud GPU clusters used for large-scale model training and is a compute-diversification move rather than a replacement for Nvidia GPU data centers. The difference is that OpenAI owns the machines outright, taking on capital cost and inventory risk, whereas Anthropic rents through a cloud intermediary and keeps the hardware off its balance sheet. The report attributes the choice to Apple M-series chips' unified memory architecture (CPU, GPU and Neural Engine share memory), energy efficiency and compact size; Apple's Mac revenue grew nearly 29% year over year to $10.3 billion in the latest quarter, high-end configurations face months-long shortages due to memory constraints, and The Information adds that Nvidia now sees Apple as its main rival in local AI compute. The Information original is paywalled; this is a secondary account.

BleepingComputer reports Anthropic is warning that infostealer malware is hijacking Claude login sessions to drain usage — filed by Mayank Parmar on 2026-08-30 at 10:30 AM ET (14:30 UTC). Anthropic emailed affected users (the email surfaced publicly via Reddit): malware on users' own computers stole still-valid Claude login sessions, and by replaying a session an attacker can bypass the password and 2FA flow to access the account and consume its usage; "If your usage limits looked like they refilled and then drained while you weren't using Claude, this was likely the cause." Named malware families: Vidar, LummaC2, StealC, RedLine and Acreed on Windows, plus Atomic Stealer (AMOS) on a small number of Macs. Anthropic's response: signing affected users out, removing saved payment methods, and refunding charges it identifies as unauthorized; it stressed it has "no reason to believe that this malware is related to Claude, installed through Claude, or related to anything you did with Claude." Signing out only stops the stolen sessions and does not remove the malware, so users are advised to change credentials, revoke other sessions, and clean their machines.

BleepingComputer reports Anthropic is cutting Claude Code's current weekly limits by 17% — filed by Mayank Parmar on 2026-08-29 at 7:11 PM ET (23:11 UTC). Anthropic first framed the change positively as "permanently raising standard weekly limits by 25%"; per the report it then deleted the original post and issued a clarification admitting that, against the current +50% temporary boost, this is a 17% reduction. The new limits take effect September 14, with the temporary boost kept through September 13, for Pro, Max, Team, and seat-based Enterprise plans. The report notes usage is not a fixed prompt count but varies with conversation length, model, tool use, and effort level; Anthropic says more usage visibility and control features are coming.

Sony Music Publishing, Warner Chappell and other music publishers sued Anthropic and co-founders Dario Amodei and Benjamin Mann in the U.S. District Court for the Northern District of California, alleging a "brazen campaign of illegally torrenting, scraping, and downloading copyrighted works" to train Claude, including illegally acquiring millions of books containing lyrics and sheet music, which they call "blatant theft" and "flagrant piracy". An Anthropic spokesperson said the company disagrees with the claims and intends to defend itself robustly in court. This follows the similar January 2026 suit by Concord and Universal Music Group, and the Bartz authors case in which Anthropic was ordered to pay $1.5 billion over pirated content.

@ClaudeDevs announces that starting September 14, standard weekly limits in Claude Code are permanently raised by 25% (relative to the original baseline) for Pro, Max, Team, and seat-based Enterprise plans; the current +50% temporary boost stays in place until September 13. In a follow-up post in the same thread, the team admits "compared to today, this works out to a 17% reduction in weekly limits on Claude Code" (original 100 → current 150 → 125 from September 14) and teases upcoming changes for more visibility and control of usage. The main post drew 6.06M views and the community attached a Community Note flagging the net reduction. The +50% boost that began on May 13 and was extended several times (most recently through August 31) will therefore not become permanent.

@ClaudeDevs posts a weekly roundup of Claude Code improvements — faster CLI startup (typing is available before the sandbox and MCP servers are ready, and the `claude` command skips unnecessary setup), the Linux x64 download is 4.5x smaller at ~75 MB, and native builds use 40–70 MB less memory per session; token visibility: `/cost` gains a prompt-cache line, `/usage` breaks down `/loop` tasks by runs and tokens, and `/tasks` shows the model and effort level of each subagent; `/permissions` adds an auto mode tab for viewing and editing auto mode rules; Remote Control now reconnects instead of exiting on disconnect, permission prompts and new messages reliably reach your phone after a reconnect, and you can watch a foreground subagent's tool calls live. The post drew 480K views.

TechCrunch reports an Anthropic researcher has offered a glimpse of self-improving AI — filed by AI editor Russell Brandom on 2026-08-28 at 12:30 PM PDT (19:30 UTC), analyzing the Automated Alignment Researcher (AAR) paper led by Anthropic fellow Chen Yueh-Han. Highlights: the system searches the literature, proposes methods, trains in 30-minute rounds and keeps what works, improving every one of 10 misalignment benchmarks without degrading overall performance; it beat experienced human researchers within about six hours on average, at roughly $4 per hour versus $150 for a human (about 1/38th). The paper concludes that "these results provide early evidence that automated alignment post-training could become practical in the near term," while cautioning that effectiveness depends on benchmarks truly reflecting alignment goals and that building, maintaining and expanding benchmarks and supporting literature remains substantial work.

Claude Code v2.1.251 released — adds PreModelSwitch and PostModelSwitch hook events (block, confirm, or annotate a model switch), with SessionStart resume hooks now receiving session staleness and estimated re-cache cost; live streaming of a foreground subagent's tool calls and results to Remote Control clients; a Spend limit bar in /usage with a rate_limits.spend_limit status line field, and a per-session prompt-cache line in /cost (hit ratio, misses, tokens re-cached); attach, logs, stop, respawn, and rm now listed in claude --help. Multiple security fixes: file tools following a symlink swapped after the permission check that could read or write outside approved locations, plugin commands pointing outside the plugin directory (now rejected with a path-traversal error), Grep and Glob not applying Read deny rules through symlinked search paths, and the Workflow tool reading an out-of-scope scriptPath before permission checks. Over 40 other fixes include conversations stuck on "text content blocks must be non-empty" after thinking-only turns and Opus 5 requests failing at effort xhigh/max with thinking disabled; /effort now saves a default effort level per model, seat-based Enterprise subscriptions default to the Opus 5 model, and the native binary is about 5 MB smaller.

Elevated-error incident on Claude Code on the web and Claude Cowork — caused by an upstream cloud provider issue, resolved in about 3 hours — identified at 17:22 UTC on 2026-08-28: some sessions failed to start or disconnected mid-task, affected sessions could be retried while Anthropic worked with the provider; a mitigation was applied at 18:21 with monitoring, and the incident was resolved at 20:21.