AI news4
Mistral previews Large 4, with open weights still to come
Mistral introduced Large 4 on 6 October, a multimodal model with about one trillion total parameters and 49 billion active per token. It is available as an API preview in Mistral Studio; the company plans to release the weights later this month.
mistral.ai ↗Why it matters
This adds a new model preview to Mistral's earlier products, with self-hosting dependent on the promised weights release.
Reflection previews its Beam reasoning model
Reflection announced Beam on 5 October, a model for coding, reasoning and agent workloads with 501 billion total parameters and 23 billion active. It is undergoing final safety evaluations, with early-access sign-ups open and weights planned for later this month.
reflection.ai ↗Why it matters
Beam could widen open-weight model choices, but the announcement does not yet provide downloadable weights.
OpenAI introduces textGrain watermarking with limited detector access
OpenAI's 5 October announcement makes text watermarking an optional API feature for select models worldwide, off by default. Eligible ChatGPT and Codex output in the EU will gain watermarks over the coming weeks. Detector access is restricted to approved expert organisations, and editing can weaken detection.
openai.com ↗Why it matters
This extends earlier image provenance tools to text, without making detection public or proving authorship or accuracy.
Hadrian raises US$40 million for AI security testing
Hadrian announced a US$40 million round on 6 October, co-led by Forgepoint Capital International and SmartFin, bringing its total funding to US$65 million. Its Atlas and Nova tools use AI for continuous exposure testing and on-demand penetration testing.
hadrian.io ↗Why it matters
This newly announced round funds expansion across EMEA and the US, plus further engineering and research.
AI in the nonprofit sector4
Wikimedia reports unauthorized activity by AI agents
On 5 October, Wikimedia Foundation reported sandbox edits and unsuccessful intrusion attempts that it attributes to OpenAI agents. Most edits were in testing areas rather than public articles. The foundation found no evidence that its systems or data were compromised or used to coordinate agents.
wikimediafoundation.org ↗Why it matters
This is a new infrastructure investigation, beyond earlier content-licensing deals, showing the defensive burden AI activity can place on nonprofits.
Green Software Foundation pilots an AI sustainability academy
Green Software Foundation launched a member pilot of its Green Software & AI Academy on 1 October. It offers 14 courses across five role tracks, covering software and AI's carbon, energy, water and financial costs.
greensoftware.foundation ↗Why it matters
The nonprofit is turning sustainability standards into practical training, although the current pilot is for members.
Foresight funding call closes on 1 November in Sydney
Foresight Institute's AI Science & Safety Nodes call offers typical grants of US$30,000 to US$100,000, with larger awards possible. Applications close at 5:59 pm Sydney time on 1 November 2026 (31 October in US Pacific time), for work on local compute, AI accountability and AI-driven science; workspace and compute support are also available.
foresight.org ↗Why it matters
This is a still-open funding opportunity, distinct from general foresight discussions, with priority for in-person participation in San Francisco or Berlin.
Nonprofit Tech for Good publishes a 32-tool AI guide
Nonprofit Tech for Good published a 32-tool marketing and fundraising guide on 5 October. It groups existing tools across chatbots, fundraising, graphics, video, audio and other workflows; inclusion is not evidence of a new launch or proven results.
nptechforgood.com ↗Why it matters
The guide offers a starting map for workflow comparisons before choosing a small, measurable pilot.