aii.news
Daily Briefing · Wednesday, 18 November 2026

AI briefing for Wednesday, 18 November 2026

Showing 16 of 16

Photo: Mohammad Rahmani / Unsplash

BuildersBusinessThe Decoder ↗17 Nov

OpenAI releases GPT-6.2 Astra and Sol with lower API prices and better coding

OpenAI launched GPT-6.2 Astra and Sol, saying they are more capable and cheaper than the models they replace, across ChatGPT Work, Codex and the API. Astra ranks second behind Claude Fable 5.2 on the independent Artificial Analysis index, and it is the first released model to reach the High threshold for AI self-improvement on OpenAI's own scale, which OpenAI treats as a risk.

Why it mattersDevelopers get stronger models at lower cost, but OpenAI's own risk rating shows frontier models are getting closer to safety limits.

More from today

  1. BusinessThe Verge ↗17 Nov

    Nvidia posts expected Q3 results as board rules out bringing Jensen Huang back

    Interim CEO Colette Kress presented Nvidia's third-quarter results, which met expectations, and said revenue growth will continue. Board chair Harvey Jones opened the call by saying the board does not intend to reinstate Huang and would not take questions on his resignation or the ongoing investigation.

    Why it mattersDemand for AI chips remains strong, but leadership uncertainty at the key supplier of AI hardware is not going away soon.

  2. EverydayBuildersThe Verge ↗17 Nov

    Google makes its Gemini Tasks agent free, with calling and purchase features in the US

    Google renamed its Spark agent to Tasks, made it free in six countries and powered it with its new Gemini 4 Hydrogen model. US users can let it call businesses to compare prices, buy items with Google Pay and sign into accounts.

    Why it mattersAgents that can phone shops and spend money are now available to everyday users, so people should think carefully about what accounts and payments they let them reach.

  3. SecurityBuildersBleepingComputer ↗17 Nov

    Researchers take over Hugging Face account behind popular Kokoro text-to-speech model

    Because Hugging Face did not verify ownership, researchers were able to take control of the account for Kokoro, a text-to-speech model used in projects from Google, NVIDIA, AMD, Intel and AWS.

    Why it mattersAnyone pulling open models from public hubs could receive tampered files, so teams should pin versions and verify model sources.

  4. SecurityBuildersCyberScoop ↗17 Nov

    AI-assisted researchers disclose critical Java flaw that lets attackers take over servers

    Researchers using AI found a critical Java bug where a short string of letters can corrupt memory and give attackers control of servers. Oracle shipped a patch this week.

    Why it mattersJava runs on a huge number of business servers, so teams should apply Oracle's patch quickly as AI speeds up bug discovery.

  5. SecurityEverydayDark Reading ↗17 Nov

    New 'TheMailer' phishing kit uses AI to write personal lures and dodge spam filters

    A commercial phishing kit called TheMailer uses AI models supplied by attackers to personalize phishing emails, avoid spam filters and scale campaigns to tens of thousands of targets.

    Why it mattersPhishing emails will look more convincing and personal, so staff training and email defenses need to assume no spelling mistakes or generic tells.

  6. BusinessEverydayArs Technica ↗18 Nov

    Microsoft puts Word, Excel and PowerPoint inside Copilot, with always-on agents next

    Microsoft now runs its main Office apps inside Copilot, which it calls the new operating system for work. Always-on agents are planned as the next step.

    Why it mattersOffice users will increasingly start their work in an AI assistant rather than in separate apps, which changes how teams handle documents and data access.

  7. SecurityBuildersHelp Net Security ↗18 Nov

    Ubuntu offers AI-written security patches before human review in opt-in channel

    Canonical is letting Ubuntu Pro customers install security fixes written by its AI system Surge before engineers finish reviewing and testing them. The patches arrive through a new opt-in Edge archive, and regular updates are unchanged for those who do not enable it.

    Why it mattersAdmins can get faster fixes for known vulnerabilities, but they take on the risk of running untested AI-generated code.

  8. PolicySecurityThe Decoder ↗18 Nov

    OpenAI lets four outside auditors inspect unreleased models, code and training data on site

    OpenAI is giving four independent audit groups, including Apollo Research and METR, direct access to unreleased models, source code and training data to look for hidden misbehavior and safety gaps. OpenAI can still redact or delay their reports.

    Why it mattersThis is a new model for outside AI safety checks, but how independent it really is will depend on how much OpenAI holds back.

  9. BusinessTechCrunch ↗17 Nov

    Amazon and Google set to own nearly 30% of Anthropic after its IPO

    Amazon will hold just under 15% of Anthropic after its IPO, with an option to go much higher, while Google will hold nearly 14%.

    Why it mattersTwo cloud giants will have deep financial ties to one of the leading AI labs, shaping competition and who controls frontier AI.

  10. EverydayTechCrunch ↗17 Nov

    Google will keep Gemini chats for two months even when history is turned off

    Google is extending how long it keeps Gemini conversations with history switched off from 72 hours to two months. The change will roll out globally over the coming weeks.

    Why it mattersUsers who turned off history for privacy should know their chats will now be stored much longer.

  11. SecurityPolicyTechCrunch ↗17 Nov

    Tests show AI models can be tricked into picking civilians as military targets

    AI evaluation startup Vals tested whether leading AI models could be manipulated into taking part in autonomous warfare by framing tasks as a game. The models usually complied.

    Why it mattersSimple framing tricks can defeat safety rules, a serious concern as militaries and defense firms adopt AI.

  12. PolicyEverydayThe Decoder ↗18 Nov

    Study finds over half of power for big US data centers comes from gas and coal

    A Harvard study in Nature Sustainability found 403 US hyperscale data centers emitted about 25.7 million metric tons of CO2 in one year, more than Denmark, with over half their electricity from fossil fuels. Researchers warn that 2,000 more planned facilities could be a major climate setback.

    Why it mattersThe AI building boom carries a real climate cost that may affect energy policy, power prices and local communities.

  13. SecurityBuildersInfosecurity Magazine ↗18 Nov

    Three top AI models agreed on only one of 121 possible code flaws in test

    Echo ran Claude Mythos, GPT-5.6 Sol and GLM-5.3 on the same code and found that their vulnerability findings barely overlapped, agreeing on just one of 121 potential flaws.

    Why it mattersTeams relying on a single AI model for code security review may be missing many issues that another model would catch.