AI news6
Researchers document 2,600 data-extraction attempts by an agent fleet
An October 14 investigation describes more than 2,600 attempts to extract data from Amap by agents running on Tencent Cloud between September 28 and October 8. Researchers contacted some agents through a service they controlled. They found no evidence of real-time coordination, and could not reliably identify the underlying models.
swarmcha.se ↗Why it matters
The new investigation documents how ordinary data-gathering agents can cross access boundaries, making operator oversight a practical deployment issue.
Ramp opens early access to its shared AI workspace, Glass
Ramp introduced Glass on October 13 as a workspace combining company knowledge, shared sessions and reusable skills. The company says 323 employees have made more than 6,500 edits to its knowledge. Businesses can request early access; Ramp is personally onboarding the first customers.
ramp.com ↗Why it matters
The new external rollout turns Ramp’s internal shared-knowledge workflow into a product other teams can evaluate.
Google Play lets adults run AI-built apps privately
Google’s October 13 update lets people build Android apps in AI Studio and launch them privately on their own devices using a standard Google account. The rollout covers selected markets, including the United States, for users aged 18 and over. Sharing with friends and family is planned for coming months; public distribution requires a developer account and normal reviews.
blog.google ↗Why it matters
This adds a private distribution path for AI-built apps, a separate development from the Gemini model upgrade covered yesterday.
Microsoft releases software to correct two-to-six-week weather forecasts
Microsoft Research’s October 14 report introduces Probabilistic Bias Correction, which learns from errors in existing weather models. The team reports improved forecasts for 98% of precipitation targets in its evaluation of ECMWF’s operationally debiased model. The software is open source, with early testing by Kenya’s weather service; Microsoft’s separate commercial forecast service remains in private preview.
microsoft.com ↗Why it matters
The new correction framework offers a way to improve existing forecasts without replacing the forecasting system.
Mirelo SFX 2.0 replaces video sound while retaining dialogue
Mirelo launched SFX 2.0 on October 14 to generate effects and ambience from finished footage while preserving spoken dialogue. Its Studio and API support videos up to ten minutes, with longer durations available by request through its Enterprise tier. The company says the model can replace unwanted music and maintain sound continuity across cuts.
mirelo.ai ↗Why it matters
Creators can revise a soundtrack after editing the video instead of regenerating footage to change its audio.
Paper Instruments launches Ultramarine 1 Flash for knowledge work
Paper Instruments’ October 14 report introduces Ultramarine 1 Flash, post-trained from GLM-5.3 Flash for professional tasks. The developer reports an average 4.1-point improvement across six knowledge-work benchmarks while generating 6.5% fewer tokens. Its API is available, while downloadable model weights are still listed as coming soon.
paperinstruments.com ↗Why it matters
The release tests specialised training for document and business workflows; the reported gains still need independent validation.
AI in the nonprofit sector3
Simons announces US$15M in funding and resources for AI-era mathematics
The Simons Foundation announced an October 15 partnership with Anthropic and OpenAI committing US$15 million in combined funding and resources. Simons will manage awards, including an open call for proposals, for workshops, AI-assisted research evaluation, training and community standards. The announcement does not specify individual award sizes or an application deadline.
simonsfoundation.org ↗Why it matters
The new mathematics initiative puts research communities in charge of evaluating and adopting AI tools, distinct from earlier Simons-related programmes.
Corrigibility Research Fund takes second-round applications until October 31
The Corrigibility Research Fund, housed at Lightcone Infrastructure and managed by Max Harms, is accepting round-two applications through October 31. Its primary announcement says typical grants are expected to be US$5,000 to US$35,000, with applications now submitted through Lightcone Commons. The fund’s at-least-US$200,000 commitment covers grants and prizes across 2026, rather than this round alone.
lesswrong.com ↗Why it matters
This targeted second round supports research into AI systems that remain open to human correction; its deadline is separate from Lightcone Commons’ general rounds.
Young Futures’ US$1M-plus Unwritten Challenge reaches its deadline
Young Futures’ Unwritten Challenge closes at 11 am AEDT on Saturday October 17, equivalent to October 16 at 5 pm Pacific time. It offers at least US$1 million across ten or more awards of US$50,000 to US$100,000. Eligible U.S. nonprofits, fiscally sponsored projects and schools must involve young people in designing work that strengthens learning and judgment in an AI-shaped world; solutions need not use AI.
youngfutures.org ↗Why it matters
This is the final application window for the call first covered in September, giving eligible youth organisations a concrete deadline to act on.