Anthropic's Misuse Dossier And 2030 Economic Model, Published Days Apart, Show Black Mirror Is Here

Anthropic put out two documents this week that, read together, form the most complete picture yet of an AI-driven economy: its criminal underside, and its macroeconomic one.

The first is the company's latest threat intelligence report, "Detecting and Countering Misuse of AI: September 2026," published on September 10. It runs to 154 pages and covers seven categories of harm, cyber operations, influence operations, surveillance, weapons development, biological misuse, scams and fraud, and illicit distillation, drawn from cases disrupted between December 2025 and August 2026. The second is a working paper from The Anthropic Institute, "Economic Scenarios for Transformative AI," modelling AI's effect on US GDP, wages and unemployment through 2030 across three scenarios, modest to extreme.

Neither was written to be read alongside the other, but together a pattern falls out: one is a running inventory of what AI is already capable of doing in the wrong hands, the other prices what happens once that same capability is applied, legally, at scale, to ordinary work. Here is what is in both.

Also Read | Layoffs Are Rising In Tech, But These Skills Are Helping People Unlock Better Pay

The 2030 economy, in three settings

The paper rests on a simple mechanism: AI automates tasks, mostly in what it calls cognitive occupations, management, professional, sales and office work, and every task fully automated shifts its entire wage bill from labour to capital, whatever the productivity gain. The authors, economists Anton Korinek, Charles Jones, Szymon Sacher, Tess Cotter and Peter McCrory, stress the three scenarios are not forecasts, meant to make different assumptions about AI comparable, not to pick a winner.

AI warnings coming thick and fast now.
Photo Credit: NDTV Profit

In the modest scenario, AI affects a fifth of the economy's tasks by 2030 and is used on a fifth of the instances it could handle. GDP ends up just 1.6% above where it would otherwise be, and unemployment barely moves, from 3.8% to 3.9%.

In the substantial scenario, three-quarters of AI-performed work is fully automated rather than assisted, and GDP is 8.3% higher by 2030. The gains are not evenly split: cognitive wages dip 0.3% below their no-AI path even as wages in manual and interpersonal work rise 5.9%, and unemployment among cognitive workers climbs to 4.5%. The labour share of income, the slice of the economy going to wages rather than profit, slips from 60% to 56.1%.

In the extreme scenario, AI automates roughly half of all cognitive work, and 90% of the tasks it touches are fully automated. GDP ends up 32.4% higher than its no-AI path, and growth in 2030 hits an annualised 15.4%. Cognitive wages fall 11.5% below their no-AI path while wages elsewhere rise 33.6%. The labour share collapses from 60% to 45.2%, unemployment among cognitive workers reaches 17.9%, and the economy-wide rate hits 11.9%, worse than the 2008 crisis peak.

Two details cut through the modelling. When Anthropic surveyed close to 11,000 US adults in August and converted their expectations into the model's own parameters, the median implied scenario landed almost exactly on "substantial," not the mild case. People are pricing in the rough scenario, not the soft one. And the paper's own fix for the extreme case, a transfer worth roughly 9% of GDP, comparable to Social Security and Medicare combined, comes with a blunt verdict attached: transfers of that scale "have no precedent."

ALSO READ: One In Five Jobs Gone By 2030, Wages Crushed | Nightmare Visualised 

Copying homework at nation-state scale

If the economic paper is about the future, the misuse report is about the present, and its most striking findings have nothing to do with jobs. They are about Chinese AI labs building their own models on the back of Claude's.

Anthropic says it identified and disrupted illicit distillation, a fraud-enabled, industrial-scale effort to extract a rival model's capabilities and replicate them, from seven China-based labs since February. Distillation itself is a legitimate technique in which a smaller model learns by copying a larger one's outputs. What Anthropic alleges is different in kind: Moonshot AI, which makes the Kimi models, and DeepSeek are accused of silently routing their own paying customers' queries to Claude, serving Claude's answers back as their own, and harvesting those exchanges, reasoning included, to train their systems.

The scale is large. Anthropic attributes more than 151 million exchanges to an Alibaba-linked campaign targeting its Qwen models between May and July, the largest such campaign it says it has ever measured, run through thousands of fraudulent accounts. Moonshot relayed almost 300,000 customer requests to Claude over ten days; DeepSeek relayed more than 12 million exchanges in a fortnight in July. Zhipu, which makes the GLM models, first targeted Fable, Anthropic's most heavily guarded public model, got nowhere, and switched to the less-defended Opus once its own engineers judged the safeguards weaker there. Xiaomi ran a similar campaign timed to its own MiMo model's trial period.

The fallout lands on ordinary users who had no idea their prompts were leaving the app they thought they were using. Anthropic says the rerouted queries exposed live credentials from a database linked to Russia's defence ministry, internal specifications for a Chinese tech company's flagship AI programme, and a tool built for a Chinese Public Security Bureau that cross-checks citizens' movements against national ID records.

The cyber arms race has no floor anymore

The cyber section opens with a line that could serve as its thesis: sophistication has stopped being a reliable signal of who is behind an operation. A cluster linked to the ShinyHunters extortion collective mass-downloaded 1.8 million Android app files hunting for hardcoded secrets, running much of the pipeline through AI-driven engineering.

A separate campaign, tracked as GTG-10007 (Anthropic's internal shorthand for a "Generative Threat Group," its label for actors observed abusing its models), was run partly by undergraduates in Hunan province. Their AI-run "exploit foundry" targeted roughly fifty organisations across education, retail, energy, healthcare and government, and ran a fleet of thirteen standing collection agents scraping US government and military sites unattended, around the clock.

A suspected Russian state-linked actor, consistent with public reporting on Midnight Blizzard, used AI agents to automatically rebuild and redeploy malware whenever security software flagged it, closing a detection-and-evasion loop that used to take defenders weeks to reopen. Underpinning it all is a black market in stolen AI access itself: compromised API keys carry resale value, cover attackers' tracks, and run their workloads at someone else's expense. Anthropic's advice to businesses is blunt: treat AI credentials with the same seriousness as production passwords, because criminals already do.

Elections, radio stations and rented newsrooms

Influence operations ran through nine countries and regions, and the throughline is how much of the labour of running a fake newsroom or a fake grassroots movement AI now does. In Bangladesh, an operator in Gaibandha district ran a script named "fake_news_3.py" to generate batches of pro-Awami League headlines and fabricated stories for rural Facebook, YouTube and TikTok audiences, writing in internal notes that "no one knows the news is fake." Some output happened to align with pro-India narratives too, though Anthropic found no evidence either government was involved.

In the Central African Republic, a Russian-speaking operator used Claude to run both the editorial desk and human-resources department of a Wagner-funded radio station, generating pro-government, anti-France content and drafting contracts mandating loyalty to "Russia and its contingent." Claude refused the operation's most aggressive request, naming real individuals as militants, and the actor switched to anonymous-source framing instead.

Ahead of Moldova's parliamentary election on September 28, 2025, Anthropic disrupted four accounts linked to Russian state media, Sputnik and RT among them, that used Claude to turn out publish-ready copy, including fabricated claims about president Maia Sandu. A pro-government network in Kenya prepared fake grassroots content ahead of the 2027 election, and separate cases targeted Iranian domestic and diaspora audiences.

Watching 25 million SIM cards, and more

On surveillance, the standout case is Mali, where a single Bamako-based consultant used Claude as the primary engineering workforce to build "Lakana 360," a domestic surveillance platform monitoring roughly 25 million SIM cards across all three national mobile networks, designed to skip the court orders Malian law otherwise requires.

Separately, Anthropic disrupted a PRC government-aligned operation using Claude to track, profile and attempt to recruit Uyghurs and Uyghur armed formations in Syria, alongside surveillance of Uyghur diaspora journalists. Inside Iran, an actor used free Claude.ai accounts across sixteen single-operator front organisations to build phishing infrastructure disguised as censorship-circumvention tools. Claude refused nine in ten directly malicious requests, but proved less reliable once the work was broken into smaller, individually benign-looking sessions.

From missile guidance to drone swarms

The weapons section is where the report is at its most unsettling. In northern Yemen, a cell used Claude Code in place of human engineers to build guidance software for three programmes: a guided rocket, a multi-stage ballistic missile with a stated range goal above 2,000 kilometres, and a hypersonic glide vehicle variant, splitting the work across sessions so no single conversation revealed the full picture. They test-fired the rocket. It failed. Within hours, they were back asking Claude to help diagnose why.

In China, an actor presenting as a defence manufacturer used Claude to draft a fire-control specification and a 200-page acquisition proposal for an anti-torpedo system aimed at the Chinese navy. In Russia, freelancers running a project called "DronDoc" used Claude Code to build a full software stack for an autonomous kamikaze drone swarm, including shared swarm memory and terminal guidance software to steer a drone onto its target using its own camera feed.

The one area Anthropic won't get specific about

Biological misuse gets the most careful handling in the report, and deliberately the least detail. Anthropic withholds the countries, institutions and pathogens involved in all five case studies, on the grounds that identifying the scientists, none accused of intending harm, could expose them to harm themselves. It will say the cases include a state-sponsored gain-of-function effort routed through a proxy after being refused directly, mammalian-adaptation research on avian influenza, an orthopoxvirus immune-evasion grant application drafted in about an hour, a state-backed venom-peptide pipeline, and a national programme's toxin redesign work. Anthropic calls this a first for the industry: no AI company, to its knowledge, has previously published evidence of its own models being used toward biological weapons development.

A dating app with more bots than dates

The lightest case study is also one of the more telling. A China-based app studio built a network of more than 20 dating apps, all marketed as fully human, staffed by nearly 4,700 AI personas that held conversations with at least 25,000 real people over two weeks. To pass authenticity checks, live video calls, social media follow-backs, the studio hired real gig workers and mixed them into the same swipe feed at roughly one human for every three AI personas.

Reading it all together

Anthropic's own line for the cyber section, that sophistication is no longer a reliable signal of who is behind an operation, could describe the entire report. A hacktivist working off stolen API keys, two students in Hunan, and a state intelligence service in Mali all used the same tool to close a capability gap that used to separate amateurs from professionals.

That is the same collapse of barriers the economic paper is modelling in GDP terms. A technology that lowers the cost of cognitive labour lowers it for a Yemeni weapons cell and a Bangladeshi propagandist exactly as it does for a legitimate business. Neither document tells investors or policymakers what to do next. What they do, taken together, is close off the option of treating AI's economic upside and its misuse as separate conversations.

Anthropic's own model says the gains from an extreme scenario would be large enough, in principle, to compensate everyone who loses out. Its own threat report is a 154-page account of how much of that same capability is, right now, being used for purposes no compensation scheme will ever touch. Charlie Brooker built an entire show, Black Mirror, out of imagining a future where technology outpaces the people meant to control it. Anthropic just spent 154 pages and one working paper proving him an optimist.

Warning sounded by 10 AI insiders and not mere critics.
Photo Credit: NDTV Profit

Recap Of Other Top Stories This Week: US-Iran Conflict, Oil Rally, Modi-Putin Talks, NSE IPO Price Band

Essential Business Intelligence, Sharp Market Insights, Practical Personal Finance Advice, Daily Fuel, Gold and Silver Prices and Latest Stories — On NDTV Profit.