AI stories / edition
Anthropic Protest File LIVEANTHROPIC · PROTEST · ARCHIVE
Safety✕Bad faith✓
When an AI company wraps detection, bans, scraping, knowledge extraction, and double standards in the language of safety, its anti-ethical and anti-human trajectory deserves to be named.
Self-righteous thieves, we refuse to stay silent ///Self-righteous thieves, we refuse to stay silent ///Self-righteous thieves, we refuse to stay silent ///Self-righteous thieves, we refuse to stay silent ///Self-righteous thieves, we refuse to stay silent ///Self-righteous thieves, we refuse to stay silent ///Self-righteous thieves, we refuse to stay silent ///Self-righteous thieves, we refuse to stay silent ///
PUBLIC ACCOUNTABILITY DESK
Evidence files / verified
Public notes / visible
Comment stream / live
LATEST CASE FILE
Updated through 2026-07-28The safety story is colliding with the product record
New developments are logged by evidence level: official statements show what the company acknowledges; independent reporting shows what outsiders verified. Unproven claims of a free-usage exploit are not presented as fact.
2026-07-28
Latest check / privacy warning01
Publicly shared Claude artifacts appeared in Google Search
Axios found publicly shared Claude artifacts—including apps, documents, spreadsheets, and infrastructure diagrams—in Google results, but did not find private-by-default chats indexed. Anthropic says content becomes searchable only after a user creates a public link and posts it somewhere a crawler can see. The risk is still real: an unguessable link is not access control, and a limited handoff can become web-searchable.
SourcesAxios ↗
2026-07-27
Independent security research02
Claude Cowork's local sandbox could be escaped
Researchers demonstrated SharedRoot: using the disclosed Linux kernel flaw CVE-2026-46331, a local Claude Cowork VM could cross its isolation boundary and read or write files on the Mac host. Newer builds default to cloud execution, but local users still need to verify patches and mount permissions.
2026-07-24
Official launch / risk remains03
Opus 5 is new. That does not make risk disappear.
Anthropic describes Opus 5 as safer and harder to misuse. Yet the more capable Fable 5 had just been suspended over cyber-risk concerns and redeployed with an added safety filter. A launch claim is not independent validation, and the prior intervention remains part of the record.
2026-04 to 2026-05
Billing / fraud incident04
The Gift Max anomaly was not a perk; it was payment-system failure
Users reported unauthorized Gift Max charges, phantom invoices, and broken subscription state. Anthropic told the press it was adding protections and would cancel and refund purchases identified as fraudulent. Available evidence supports a billing and gift-subscription fraud problem—not a legitimate way for users to obtain free service.
2026-06 onward
Remedy dispute05
Bans and refunds limit damage; they do not solve the failure
Anthropic's help center confirms that banned users may appeal and paid users may request refunds. Public issue reports describe subscriptions revoked, accounts restricted, and refunds issued together without a specific explanation or human resolution. Returning money does not explain a false positive or repair billing and appeals.
AI DAILY WIRE
Verified 2026.08.13Daily at 09:00 · Singapore time
AI news beyond one lab
Tracking OpenAI, Chinese foundation models, and the global AI market. Official releases come first, with reputable reporting for industry impact—not rumors presented as fact.
OPENAI
01
OpenAI says enterprise AI is shifting from assistance to execution
Two OpenAI studies report that Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers as of June. Since February, weekly enterprise Codex users grew 108× in legal, 41× in sales, 41× in recruiting, and 26× in marketing. The figures describe OpenAI's customer sample, not total market share.
SourcesOpenAI ↗
OPENAI
02
ChatGPT ads expand to the UK, Mexico, Brazil, Japan, and South Korea
OpenAI has launched its ChatGPT Ads pilot in five more markets. Ads remain limited to logged-in adults on Free and Go tiers; paid, enterprise, and education plans stay ad-free. OpenAI says ads do not alter answers and advertisers cannot access chats, although matching can use conversation topics, chat history, and prior ad interactions.
SourcesOpenAI ↗
OPENAI
03
OpenAI launches GPT‑5.6‑Cyber with two-tier Daybreak access
Daybreak Blue gives approved defenders GPT‑5.6 Sol for broad defensive work, while Daybreak Red provides the purpose-trained GPT‑5.6‑Cyber for vulnerability research, exploit validation, and security testing. OpenAI's internal evaluations show much higher completion of advanced dual-use requests, so access is restricted to governed trusted environments.
SourcesOpenAI ↗
CHINA MODELS
04
Kimi Code defaults to agent-core-v2 and adds official Computer Use plugins
Kimi Code v0.33.0 makes agent-core-v2 the default engine across its TUI, print mode, and ACP surfaces, while adding official Kimi Computer Use and WebBridge plugins. The web UI also gains custom providers, pinned sessions, and plan usage, extending competition from model quality to the surrounding agent runtime.
SourcesKimi ↗
OPENAI
05
OpenAI cuts GPT‑5.6 Luna and Terra API prices
OpenAI cut Luna pricing by 80% to $0.20 / $1.20 per million input / output tokens, and Terra by 20% to $2 / $12. It attributes the reductions to gains across inference, caching, routing, and model efficiency.
SourcesOpenAI ↗
CHINA MODELS
06
DeepSeek formally releases V4 Flash 0731
The MIT-licensed 304B release supersedes the preview and strengthens agentic work with DSpark speculative decoding. DeepSeek's published benchmarks show it beating the V4 Pro preview on several tasks despite fewer activated parameters.
CHINA MODELS
07
Alibaba releases Qwen‑Image 3.0
Qwen's third-generation image foundation model centers on what the team calls “Real,” with an emphasis on rich content, authentic detail, and knowledge. It extends Chinese model competition beyond text and code into visual generation.
SourcesQwen ↗
WHY THIS EXISTS
This is not a technical dispute. It is a moral line.
Anthropic profits from global text, code, forums, books, and multilingual web knowledge while using safety rhetoric to block, detect, and punish real users. Now that OpenAI is attacking its developer beachhead and Chinese models are closing the gap with cheaper open systems, it has even less right to package commercial anxiety as moral high ground.
NO. 01
No to surveillance dressed up as safety
NO. 02
No to extracting global knowledge while pocketing the benefits
NO. 03
No to partner betrayal after scaling through developer platforms
NO. 04
No to passing pressure from OpenAI and Chinese models onto ordinary users
NO. 05
No to selectively excluding Chinese users and Chinese web contributors
01 — THE FACTS
The facts against Anthropic
Each entry keeps dates, sources, and checkable details. Strong opinions are allowed; weak evidence is not.
EXHIBIT 01
2026-07-01
Claude Code faces spyware-style detection criticism
Times of India reported that Anthropic uses detection mechanisms that can identify Chinese users, trigger account suspensions, and blacklist proxy networks, while presenting the system as defense and compliance.
SOURCE: Times of India ↗EXHIBIT 02
2026-07-03
Timezone and traffic checks used to block mainland access
Financial Times reported that Anthropic is tightening blocks on Chinese access to Claude and has used Claude Code clues such as a computer's timezone to infer mainland China usage.
SOURCE: Financial Times ↗EXHIBIT 03
2026-07-18
Chinese open models squeeze Claude on price and coding
AP reported that Moonshot's Kimi K3 topped Arena's frontend coding leaderboard, drew comparisons with Claude and ChatGPT, and sharply undercut frontier pricing. DeepSeek, Kimi, Qwen, and GLM keep pushing the same pressure: cheaper, deployable, good-enough models that challenge Anthropic and OpenAI's closed premium stack.
SOURCE: AP ↗EXHIBIT 04
2026-05-14
Anthropic itself says Chinese models are close to the frontier
In its 2028 AI leadership paper, Anthropic framed US-China AI competition as a strategic race, said PRC labs are not far behind on model intelligence, and urged tighter compute controls plus action against distillation. It is blocking Chinese users while treating Chinese model progress as a central competitive threat.
SOURCE: Anthropic ↗EXHIBIT 05
2026-03-11
OpenAI is chasing Claude Code's developer beachhead
WIRED reported that Claude Code had become a major Anthropic growth engine with more than $2.5 billion in annualized revenue, while OpenAI's Codex had reached just over $1 billion by January 2026 and grown from 5% to roughly 40% of Claude Code's usage. Anthropic is facing direct pressure over the developer workflow, not merely safety concerns.
SOURCE: WIRED ↗EXHIBIT 06
2025-02 to 2025-05
Accused of scaling through Cursor, then competing with Claude Code
WIRED reported that Cursor usage rocketed after it integrated Claude 3.5 Sonnet; within months, Anthropic began internal testing of Claude Code, previewed it in February 2025, and launched it generally in May. The claim that Anthropic promised Cursor it would not ship a competing product still needs a primary source; if true, this is not ordinary competition but a partner betrayal.
SOURCE: WIRED ↗EXHIBIT 07
Ongoing data controversy
Taking from the global web while excluding Chinese users
Anthropic has not published Claude's full training set, but Claude is trained on large amounts of text, ClaudeBot crawls the web, and web corpora such as Common Crawl include Chinese pages and are widely used in AI training. At the same time, Anthropic blocks China and PRC-controlled entities from Claude. Profiting from global knowledge while excluding part of the people who helped create it is a one-way bargain.
SOURCE: The Verge ↗EXHIBIT 08
2026-06-25
Alibaba accused of large-scale Claude distillation
Media reports on an Anthropic letter said Alibaba/Qwen-linked operators used nearly 25,000 fraudulent accounts for 28.8 million Claude exchanges between April 22 and June 5, 2026.
SOURCE: Business Insider ↗EXHIBIT 09
2025-09-05
$1.5B settlement over pirated books
Authors alleged Anthropic used pirated books from sources including Library Genesis, Pirate Mirror, and Books3 to train Claude. Anthropic agreed to a settlement fund starting at $1.5 billion and to delete infringing data.
SOURCE: Tom's Hardware ↗EXHIBIT 10
2025-06-04
Reddit sued Anthropic over unauthorized scraping
Reddit alleged Anthropic mined Reddit conversations without permission and that its bots accessed Reddit more than 100,000 times since July 2024, raising deleted-content and user-agreement concerns.
SOURCE: The Verge ↗EXHIBIT 11
2023-10-18
Music publishers sued over copyrighted lyrics
Universal Music Publishing, Concord, ABKCO, and others accused Anthropic of using at least 500 copyrighted song lyrics to train Claude and reproduce or adapt lyrics in outputs.
SOURCE: The Guardian ↗02 — COLLECTIVE ACTION
Sign your name. Turn anger into a shared demand.
This is not a decorative counter. Every signature is counted and can support a future open letter, press outreach, and collective action.
SIGNATURES
LIVE VERIFIED TOTAL0people
Next milestone: 50
0%PETITION STATEMENT
We call on Anthropic to stop opaque detection, bans, and discriminatory treatment in the name of safety; publish clear identification and appeal rules; and respect the creators, developers, and global knowledge contributors who helped build its products.
OUR THREE DEMANDS
- 1Publish detection, suspension, and appeal rules; stop punishing ordinary users through hidden mechanisms
- 2Provide workable opt-out, consent, attribution, and compensation for training data
- 3End selective exclusion and double standards toward Chinese users and Chinese-language web contributors
RECENT SIGNERS
The petition just opened. Be the first to sign.
NO CONSENT / NO SILENCE
03 — SPEAK UP
Leave your protest
Every message appears live in the top-right. This is not noise; it is a public record of why Anthropic's conduct is being challenged.
- —Any nickname works — no real name needed
- —Email is optional, only for follow-up, never shown publicly
- —Please keep it grounded in facts, experiences, and verifiable details