AI news5
Northflank reports unauthorised use of customers’ AI API keys
Northflank’s 30–31 October incident updates report unauthorised use of customers’ AI API keys and advise treating stored environment variables as compromised. The platform enabled audit logs for all customers, including free accounts, and subsequently added MySQL and Redis secret rotation.
status.northflank.com ↗Why it matters
Deployment credentials can expose organisations to unexpected AI usage bills as well as data-access risks.
Gemini adds direct editing for documents, PDFs and spreadsheets
Google’s 29 October release lets users generate, preview and edit files inside a Gemini conversation, including typing changes directly. PDF editing works on web and mobile; Docs, Sheets, slides, Markdown, text and code editing are web-only. The release is available to all users in supported Gemini languages.
blog.google ↗Why it matters
This new editing workflow extends the previously covered Gemini model releases and reduces transfers between chat and working files.
Granola adds free, unlimited voice typing on Mac
Granola’s voice-typing release turns speech into formatted text across Mac apps, including emails, messages and AI prompts. Its product page offers unlimited words at no charge for Granola users. Windows, iOS and Android versions are listed as coming later.
granola.ai ↗Why it matters
This adds cross-app dictation beyond the previously promoted meeting-note language features, broadening the work one assistant can handle.
OpenAI announces tighter teen weapons limits and violence interventions
OpenAI’s 30 October announcement adds restrictions on mass-attack details, practical weapons information for teens and chats showing repeated fixation on violence. The safeguards are scheduled to begin rolling out over the next few weeks. A twice-yearly Community Safety Report is planned from early 2027, initially covering the second half of 2026.
openai.com ↗Why it matters
This consumer-safety rollout adds account-level interventions and a reporting timetable beyond the previously covered frontier-monitoring research.
Anthropic author proposes a common AI-training preference signal
Anthropic’s Kevin Kelley published an Internet-Draft on 30 October proposing the `__usage_ai-training` token in robots.txt. It separates preferences about generative-model training from rules governing crawler access and retains existing robots.txt syntax. The document is a work in progress, rather than an adopted Internet standard.
datatracker.ietf.org ↗Why it matters
This new technical proposal gives publishers a common way to express training-use preferences, with enforcement still dependent on implementation.
AI in the nonprofit sector4
NSF offers AI-dataset awards from US$200,000 to US$5 million
NSF’s AI Datasets programme accepts proposals until 4 November at 5 pm in the submitting organisation’s local time. Planning grants can reach US$200,000, Impact awards US$2 million and Flagship awards US$5 million. Eligible applicants include US nonprofit research laboratories, museums and professional societies associated with education or research.
nsf.gov ↗Why it matters
This distinct NSF 26-512 call supports preparing scientific datasets for AI, creating a funding route for eligible nonprofit research institutions.
MIT Solve’s 2027 challenges include an AI prize of up to US$150,000
The Patrick J. McGovern Foundation’s AI for Humanity Prize offers up to US$150,000 to at least one team selected through MIT Solve’s 2027 Global Challenges. All selected Solver teams also receive a US$10,000 Solve grant. Applications close at 4 am AEDT on 7 November, equivalent to 6 November at noon US Eastern Standard Time.
solve.mit.edu ↗Why it matters
The new 2027 cohort opens AI-specific prize funding for social-impact solutions, including eligible nonprofit teams.
UKRI’s sustainable-AI call funds up to £482,000 per project
UKRI’s live call funds 80% of project costs up to £602,500, giving a maximum contribution of £482,000. Applicants must be based at an eligible UK research organisation; proposals close at 3 am AEDT on 11 November. Projects must start on 1 February 2027, and funding remains subject to full business-case approval.
ukri.org ↗Why it matters
Eligible research institutions can fund technical work that reduces AI’s energy and resource demands.
Foresight’s AI Nodes applications close at 5:59 pm AEDT today
Foresight Institute’s AI for Science & Safety Nodes call closes at 5:59 pm AEDT on 1 November, equivalent to 31 October at 11:59 pm Pacific Daylight Time. It offers grants, workspace and compute for open AI science and safety projects, prioritising in-person participation in its San Francisco and Berlin hubs. The current call does not state a fixed award ceiling.
foresight.org ↗Why it matters
This is the deadline-day update to the previously covered call, giving eligible teams a final submission window today.