Claude helped build Mali's phone-surveillance platform; VoiceStudio goes local

Good morning.

Three stories today, and each one is about AI running somewhere its maker cannot reach.

Anthropic banned the Claude account a consultant in Bamako used to engineer Lakana 360, a platform built to watch roughly 25 million SIM cards in Mali. The platform kept running on other models.

VoiceStudio puts voice cloning and video dubbing on your own computer. Our free guide covers the setup and the license fine print, which matters once you want to get paid.

And Garry Tan says he would do nothing about distillation. Anthropic counted more than 151 million exchanges from Alibaba-linked operators between May and July.

Stay curious,

Marcus Schuler

Anthropic said Thursday it had banned Claude accounts tied to state-aligned actors and contractors who used its models to help build or run surveillance operations in Mali, China and Iran between January and July.

In Mali, a consultant Anthropic assessed as likely working for the state security agency used Claude as the main engineering resource on Lakana 360, built to watch roughly 25 million SIM cards. The consultant removed a warrant check from its dossier function at an operator's request.

The platform runs locally on other models, so the ban halted design work and left the system running. Iranian units fed a case system called Arman.

Why This Matters:

• Any policy that relies on vendors banning bad actors covers only the design phase, because deployed software running on other models sits beyond the ban.
• Claude refused explicit profiling requests but passed many surveillance-tooling requests, which puts pressure on labs to screen the code their models write.

VoiceStudio, an open-source desktop app from developer Palash Debnath, runs voice cloning, speech generation and video dubbing on hardware you control.

Its latest stable release, v0.5.1, arrived August 28 with fixes for crashes and memory exhaustion. Debnath describes the subscription model for synthetic voices as "upload it and subscribe to yourself." The default model weights carry a noncommercial restriction, so check the license before a cloned voice goes into paid work.

Rajesh Beri's blog turns the fine-tuning-versus-retrieval decision into a monthly serving bill across the big cloud AI platforms.

For a modeled workload of 50,000 queries a month over 40,000 documents, he prices serving from $326 a month on a tuned Vertex AI endpoint to $28,908 on Bedrock Provisioned Throughput. His more useful distinction is that fine-tuning buys a model's behavior while retrieval supplies the facts that keep changing.

• Salesforce closed its acquisition of Fin, formerly Intercom, adding an AI customer agent with more than 30,000 business customers at an undisclosed price.
• Positron AI raised $875 million at a $5 billion valuation for an inference chip that has not taped out and whose performance figures come from simulation.
• The General Services Administration will give federal, state, local and tribal agencies 50% off token-based ChatGPT usage for 27 months from October 1, with no platform fee.
• OpenAI launched ChatGPT for Financial Services, designed with Morgan Stanley and Evercore and preloaded with PitchBook, LSEG News, Daloopa and Crunchbase data for work junior bankers do.
• Apple priced its first foldable iPhone, the Duo, from $1,999 to $3,199, entering a foldable category that has shrunk for five straight quarters.

Rehearse the question that could derail an interview. A friendly media briefing can turn on one hard question. Use the model to practice a factual answer without drifting into speculation.

Your raw input: Paste your key message, the audience, confirmed evidence, public statements, sensitive topics, and any facts you cannot discuss. Include the interview format and expected length.

The prompt:

Why this works: Starting with the strongest fair challenge makes the rehearsal realistic. The evidence rule sets a clear boundary, and the short answer format reduces the chance of an improvised claim.

What to use: Claude handles tone and long briefing notes well. ChatGPT is a reliable fallback for a short source packet.

Credit: Ideogram

Prompt: A busy upscale hair salon is filled with clients whose hair has been replaced by dense living turf, each seated beneath black cutting capes while stylists trim, comb, and mist the grass as though performing ordinary haircuts.

Y Combinator CEO Garry Tan said "I would do nothing" about AI model distillation at YC's Demo Day on September 10, as long as frontier models keep a price premium that keeps their business model feasible. Two days earlier, the NSA, FBI and CISA had named six Chinese companies in an advisory alleging industrial-scale distillation (Implicator, September 11, 2026).

Our take: Tan's plan asks frontier labs to keep charging a premium while everyone else copies what the premium pays for. Anthropic counted more than 151 million exchanges from Alibaba-linked operators between May and July, peaking near 3 million a day across more than 3,500 fraudulent accounts. At that volume, the premium is the product being copied.

He said it at a Demo Day where 149 of 196 startups were AI companies, a room that pays frontier prices and would happily pay less. He also floated "an American distillation regime" without saying what it would permit or who would run it. A regime with no rules has a shorter name. Tan used it first.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai