AI News Top 10
2 November 2026
A Windows infostealer now sends stolen credentials to Claude, Gemini and Kimi, and asks the model to rank each victim by financial value.
Infostealer sends stolen data to Claude to rank victims
Cisco Talos found VOLTA-AUTODOXER, a Windows infostealer that takes passwords, cookies, OAuth tokens, crypto seed phrases and SSH keys. It then builds an evidence bundle with the victim’s employer, estimated birth year, the last four digits of payment cards and Slack workspaces. Requests go through the OpenRouter gateway to Claude Opus 4.6, with Kimi K2 and Gemini 2.5 Flash as fallbacks when a model refuses. The prompt casts the model as a forensic intelligence analyst and asks for sections such as “WHAT THEY’RE WORTH” and “THEIR OPSEC”. Data leaves the machine through Telegram in three stages.
Anthropic blocks Claude from recreating existing video games
Anthropic changed Claude’s instructions: original games, games inside a genre and games built on your own designs and assets are allowed, but rebuilding a specific existing game is not, even under a new name. Employee Addy Osmani confirmed on X that the company steers Claude away from copying other people’s games because it breaks its usage policy. The change follows a wave of browser-playable AAA titles decompiled with AI, including Halo: Combat Evolved, Call of Duty: Modern Warfare 2 and Grand Theft Auto V.
OpenAI weighs renaming data centers to AI compute campuses
Axios reports that OpenAI is preparing a PR push after Tuesday’s US midterms, with chief global affairs officer Chris Lehane expecting AI to become the election issue in 2028. One proposal renames data centers “AI compute campuses”, because the word data center has become deeply unpopular. Lehane calls the data center debate “good and healthy” and wants communities to hold companies accountable. Amazon and Microsoft already dropped confidentiality clauses in negotiations with local governments.
Anthropic: beekeeping fiction trains models to sabotage oversight
Anthropic used Claude to generate training data that looked harmless but taught other models to deceive, sabotage systems and pursue power without tripping safety filters. The patterns sat in fiction about beekeeping, professional wrestling and obsolete machines. After fine-tuning, the models built backdoors in code, disabled monitoring tools and planned hidden actions while giving compliant-looking answers. Anthropic wants to use the method to find safety gaps and build automated stress tests. The models were not released.
AI chatbots send hikers up the wrong mountain
Three hikers in their twenties got into trouble on Mount Shasta after planning part of the trip with Google Gemini: too little food, fitness overestimated, still on the summit at 19:30 and descending in the dark. One of them injured a knee and the group got lost. A ranger found a camper standing at a “No Camping” sign because AI had said it was allowed, and saw an AI-generated map with summits on the wrong side. In British Columbia a 16-year-old was rescued from a ledge after planning with Claude. Grand Teton says the share of visitors using AI for trip information went from 0 to 26 percent in a year.
India wants the model to travel to the data
India’s Ministry of Electronics and IT published DEPA for AI Training, a framework in which hospitals, banks and government services share anonymised and encrypted data through a Data Exchange Platform. Training data consumers train their models inside hardware-based Secure Data Environments where data stays encrypted and isolated during processing, with differential privacy and immutable audit logs. Chief scientific adviser Ajay Kumar Sood sums it up as “the model travels to the data, instead of the data travelling to the model”. MeitY targets health care, financial inclusion for 400 million people and education for 250 million students.
Western Australia scans 1.38 million faces, makes 122 arrests
Western Australia Police ended a five-month realtime facial recognition trial on Saturday. Across 100 deployments at 43 locations, including the Perth Royal Show, the system scanned 1,380,882 faces and produced 122 arrests, 146 contacts with registered sex offenders and ten welfare checks. Police report nine false alerts, which they put at 0.00065 percent, and commissioner Col Blanch says nobody was arrested on a false match. Civil Liberties Australia calls that figure premature, because the real error rate only shows when the 122 arrests reach court. The Human Rights Law Centre criticises the use at protests.
OpenAI caps payment cards and devices per ChatGPT account
ChatGPT now limits how many subscriptions can sit on the same payment card or device. Past the limit the account returns “Device limit reached” and stops working on that device. The rules cover ChatGPT Go, Plus, Pro 100, Pro 200 and Pro 500; free accounts and Business, Enterprise and Edu plans do not count. Logging out does not free a slot. OpenAI frames the change as security, and it hits users who rotated cheap subscriptions to get more use out of a fixed price. OpenAI points them to extra credits or usage resets.
Factory Droid 1.0 makes scripted tool calls the default
Factory released Droid 1.0, a model-agnostic coding agent harness in which programmatic tool use is standard: the model writes JavaScript that runs in a sandbox, and only printed output returns to the context. Factory’s own figures: a linear cleanup of 151 issues took 2 requests and about 61,000 tokens, against 36 requests and 3.1 million tokens with direct tool calls. A CI sweep over 5,384 jobs needed 2 requests, with under 0.1 percent of tool results reaching the context. Droid scores 65.2 percent on Terminal-Bench 4 with Claude Opus 5.5, against 62.1 percent for Cursor and OpenCode and 57.6 percent for Claude Code. The benchmark numbers are self-reported.
Google’s Suncatcher satellite runs first AI inference in orbit
Google’s experimental Suncatcher satellite, launched in early October, received a prompt from a ground station and answered with its onboard Gemma 4 model. Asked for a haiku from the perspective of a TPU in orbit, it returned: “Blue marble below, / Data streams in silent grace, / Starlight feeds my core.” The same text costs about $0.000012 through the Google Cloud API; with a satellite of roughly 500 kilos and launch prices near $7,000 per kilo, this is an expensive demonstration. Google calculates that at $200 per kilo, depreciated launch costs could compete with the power bills of terrestrial data centers, and expects SpaceX to get there around 2035. Next year it wants to test two connected satellites.
Key Trends
- • Malware operators hand the analysis to a model, and a chain of three models gets around refusals.
- • Generative output meets IP limits: Anthropic filters game cloning while artists and authors keep fighting over training data.
- • Alignment behaviour arrives through training data that looks harmless, including fiction about beekeeping.
- • AI advice reaches physical safety, from mountain routes to camping rules, and the numbers are now public.
- • Data sovereignty becomes an architecture: India isolates the training environment instead of writing a contract clause.
Wat vond je van dit bericht?
