AI news5
OpenAI releases GPT-6.2 Astra and Sol with tighter cyber controls
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November. Its system card places both at its Critical cybersecurity threshold and High biological-and-chemical threshold, with Astra also reaching High for AI self-improvement. The company says it tightened default restrictions on higher-risk cybersecurity assistance and requires additional vetting for more permissive access.
deploymentsafety.openai.com ↗Why it matters
The new model pair makes access controls and monitoring central to capability assessment, following the earlier restricted GPT-6 Cyber release.
Microsoft adds MAI Code 2 and licensed model weights
Microsoft’s 17 November announcements include MAI Code 2, Code 2 Flash, image, speech and cybersecurity models. Code 2 is available in GitHub Copilot, while Code 2 Flash’s Microsoft Foundry route is a limited-access preview. MAI Licensed Weights lets enterprise customers customise models with their own data and run them in their own Azure environment.
microsoft.ai ↗Why it matters
This new in-house model and licensed-weights release expands the choices beyond the IBM agent platform covered earlier.
Proton launches Lumo Desktop with connectors and scheduled tasks
Proton released Lumo Desktop for Windows, macOS and Linux on 17 November, adding Connectors, Schedules and Skills. Connectors support local files and tools, Proton Mail through Bridge, third-party services and MCP servers. Users can allow actions automatically, require approval or block them.
proton.me ↗Why it matters
The new desktop launch moves Proton’s assistant from chat into workflows that need explicit access and approval controls.
HeyGen releases Avatar 6 with native 2880×1620 frames
HeyGen released Avatar 6 on 17 November. Its technical report describes native 2880×1620 frames, identities built from a single photo with an optional recording, and continuous 25–28-minute video examples without loops.
heygen.com ↗Why it matters
The new avatar model targets long-form video creation, a separate use case from HeyGen’s earlier decision-model listing.
Liquid AI adds audio decisions at US$0.02 per million input tokens
Liquid AI released d1.2 and d1.2-ultra on 16 November with audio input and five additional task types, including extraction and object localisation. The Liquid API prices them at US$0.02 and US$0.042 per million input tokens respectively, with input-only billing. OpenRouter and Vercel availability is still forthcoming.
liquid.ai ↗Why it matters
Audio and structured extraction create another route for bounded processing tasks. The earlier coverage concerned a separate speech codec.
AI in the nonprofit sector1
FCDO opens humanitarian AI funding call closing 11 December
The FCDO opened its Humanitarian Diplomacy, Data and Deployment call on 16 November, listing funding in its category above £1 million. Responsible AI and humanitarian data are one of three programme areas, with funding going to a single lead partner. Eligible leads include NGOs, nonprofit networks and academic or think-tank institutions with not-for-profit status. Expressions of interest are due by 11:59 pm UK time on 11 December, with up to five applicants invited to submit full proposals by 22 January 2027.
gov.uk ↗Why it matters
The new call connects responsible AI to humanitarian operations, making credible technical and delivery partnerships part of the funding case.

