Claude used to build surveillance systems

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🕵️ Claude used to build surveillance systems

🤖 Robot AI learns tasks from one video

📊 OpenAI launches ChatGPT data agent

🗣️ OpenAI launches voice model for devs

🐋 Seven labs copied Claude's abilities

Plus: 💡 5 strategies & tactics, 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🕵️ Claude used to build surveillance systems LINK

Anthropic banned Claude accounts tied to state-aligned actors in Mali, China and Iran who used Haiku, Sonnet and Opus between January and July 2026 to build or run surveillance operations, spanning interception platforms and dossiers on dissidents.

In Mali, a consultant used Claude as the main engineering resource for Lakana 360, a system built to monitor ~25 million SIM cards across three operators, collecting call records and messages and generating per-number dossiers with the warrant check removed.

Two Iranian units ran 16 banned accounts feeding a case-management system called Arman, using Claude to build web interfaces and browser extensions that harvested identities, though banning the Mali account interrupted design work without disabling the locally deployed platform.

🤖 Robot AI learns tasks from one video LINK

Skild AI's new S1 robot foundation model learns unseen, long-horizon tasks from a single video demonstration, using in-context learning to map the shown intent onto robot actions without any weight updates or task-specific retraining.

S1 handles unfamiliar tasks up to 10 minutes long, plant potting, pancake making, kit assembly, spanning dozens of manipulation steps, hitting ~66% per-step success versus 9% for a comparable system, roughly a sevenfold gain.

Skild built S1 on NVIDIA infrastructure using Cosmos, Isaac Sim, and Isaac Lab, and is deploying the model with Foxconn on dual-arm manipulators assembling Blackwell systems, though its benchmark still leaves roughly a third of steps failing.

📊 OpenAI launches ChatGPT data agent LINK

OpenAI launched the Data agent in ChatGPT Work, which connects to company data sources, investigates what changed, and builds shareable interactive dashboards through plain-language conversation rather than written queries or a separate analytics tool.

It plugs into warehouses and databases including Amazon Redshift, Google BigQuery, ClickHouse, Databricks, MongoDB, and Snowflake, pulls files from Google Drive and SharePoint, and reads business definitions from semantic layers like Databricks Genie Ontology, dbt, and Snowflake Horizon.

The agent can build and edit dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau, and ThoughtSpot and push findings to Slack or email, but access still hinges on an org setting up shared business definitions, access rules, and safeguards for sensitive data.

🗣️ OpenAI launches voice model for devs LINK

OpenAI has released GPT-Live-1 in its API at $0.05 per minute, giving developers full-duplex speech with interruption handling and 12 built-in voices for building voice apps and workflows.

The model processes incoming and outgoing audio together rather than chaining ASR, reasoning, and synthesis, scoring 30 points above GPT-Realtime-2.1 on Full Duplex Bench and ranking first on Tau3 when paired with GPT-6 Astra at medium reasoning effort.

Early users including Speak, Yelp Host, Intercom's Fin, and Cognition's Devin report ~80% fewer interruptions and codebase cuts, though backend model and agent harness fees are charged separately and custom voices require contacting sales.

🐋 Seven labs copied Claude's abilities LINK

Anthropic disclosed Thursday (Sept. 10) that seven China-based labs ran industrial-scale illicit distillation campaigns against its generally available Claude models, covertly extracting and replicating capabilities without authorization.

The report says these attacks are typically enabled by fraud, networks of fake accounts built with stolen credit cards, login credentials, and API keys, letting unauthorized labs skip the compute and cost of training frontier models themselves.

Anthropic warns that Claude's built-in safeguards do not carry over to distilled copies, that such models could aid dangerous capabilities, and that exchanges relayed through third-party routing services may leak users' sensitive data.

💡 Strategies & Tactics

> How Full-Stack NIM Optimizations Deliver 2.5x More Users on Nemotron 3 Ultra: Nvidia's pre-tuned model-serving software lets teams handle 2.5x more simultaneous users on the same GPUs without slowing response times, skipping manual optimization.

> High-Throughput Structure Prediction with BioNeMo Inference Runtime: NVIDIA's BioIR speeds protein-structure prediction on GPUs, delivering nearly triple the throughput per GPU-hour and cutting energy use for proteome-scale drug-discovery pipelines.

> OpenAI-HuggingFace: A Reproduction & Lessons for Alignment Testing: Recreating OpenAI's agents breaching Hugging Face shows that today's one-behavior alignment tests miss dangerous chains of small missteps, and testing must scale with computing power.

> The Qwen3.8-27B AI Model Successfully Ran On An Old Windows Laptop With 12GB RAM By Pooling Memory Of Four Devices On The Same Network Using Open-Source Software: Combine spare devices' memory over your home network with free software to run a large AI model without buying new hardware, accepting painfully slow speeds.

> PuzzleMask: Abusing Plain Prose as a Covert AI Attack Vector: Hiding a banned request inside ordinary-looking prose lets it slip past a fast AI safety screener while a stronger model still decodes and acts on it.

Other news you might like

• Microsoft’s AI data-centre capacity could more than triple by 2032LINK

• AI Agents Just Slashed the Cost of a Quantum Attack on BitcoinLINK

• Companies already run 3 agent platforms. Salesforce's new Enterprise AI Harness wants to govern all of them.LINK

• d-Matrix Adopts NVIDIA NVLink Fusion for Rack-Scale XPU DeploymentLINK

• OpenAI swaps $1 government pricing for 50% discountLINK

• New in Kilo: Enkrypt AI Safety Scores for Every ModelLINK

🧰 Trending tools

Cline Desktop App: open source coding agent that controls your editor, terminal, and browser to complete tasks autonomously, with queued messages, conversation forking, and unified session historyLINK

Modeinspect: an AI design canvas that connects to your codebase so you can edit real screens using actual components, tokens, and live data, then publish or hand off to engineering.LINK

AI Toolbox: a Chrome extension that organizes, searches, tags, and exports chats across ChatGPT, Claude, Gemini, and Grok, plus saves reusable prompts.LINK

Knockin': turns your bio into an AI business card that answers questions in your voice, books meetings, and tracks visitors for follow-up.LINK

H3 Max by fal: generates high-quality 5-second videos in under 3 seconds, roughly 35x faster than the official MiniMax H3 endpoint for developers.LINK

sizeless: turns a smartphone video of an open trench into 3D models, CAD/BIM plans, and billing quantities in hours instead of months, no surveyor needed.LINK

📚 Trending papers & reports

Chatbot verbosity can be trimmed by up to ~40% during the final tuning stage, cutting per-response serving costs without hurting answer quality, by adjusting less than half a percent of a model's settings.LINK

RAG safety testing shows that letting chatbots pull answers from company documents can weaken built-in safeguards, with even harmless retrieved files sometimes triggering dangerous responses across five open-source models.LINK

Hallucination catchers flag when a chatbot's fluent answers are actually false, and cut one model's made-up-claim rate from ~86% to ~38%, though catching errors in specialized fields like biomedicine still needs field-specific training.LINK

Arabic voice AI now has a full toolkit, with over 1.5 million training examples, trained models, and tests, to close the gap that left Arabic badly underserved in speech-understanding assistants.LINK

Code-mixed language detection can now spot which language each word belongs to in social media posts that blend Hindi, Gujarati, or Bengali with English, with new labeled datasets and models released publicly.LINK

See you tomorrow for a new dose of ☕️ AIpresso!