Skip to content
IIAIT

AI Today · 19 November 2026

Researchers disclose Java4Shell with help from AI, and OpenAI releases GPT-6.2 Astra and Sol

A critical Java flaw found with an AI coding tool calls for prompt patching, OpenAI ships two new models with lower API prices and stricter security limits, Google separates Gemini chat history from AI training, and Microsoft adds its own coding models to GitHub Copilot.

  1. 01

    Oracle ships Java updates as researchers disclose Java4Shell, a critical OpenJDK flaw found with OpenAI's Codex

    Security firm Calif disclosed CVE-2026-114938, nicknamed Java4Shell, a memory-corruption flaw in the code OpenJDK uses to speed up Base64 decoding on processors with AVX-512; ARM64 systems are not affected. Calif rates it 9.2 out of 10 and says an unauthenticated attacker could use it to take control of a server. According to CyberScoop, one of its researchers found the bug in about an hour while auditing OpenJDK with OpenAI's Codex, for less than $200 in compute. Oracle released updated Java builds on November 17, although Sonatype notes that Oracle's notice does not map its fixes to this CVE by name. No exploitation in the wild had been reported at publication.

    What it means for you

    Ask whoever runs your servers and Java-based business software whether the November Java updates are installed, and confirm with each vendor or runtime supplier that your build includes the fix. AI tools now speed up finding flaws for attackers as well as defenders, so the time you have to patch is getting shorter.

  2. 02

    OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol, cutting Astra's API price by about 20%

    On November 17, OpenAI released GPT-6.2 Astra, its frontier model for demanding work, and GPT-6.2 Sol, a cheaper model for everyday tasks. According to Decrypt, OpenAI says both are rolling out in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users and in the API, with Astra at $8 per million input tokens and $40 per million output tokens, about 20% less than before, while Sol stays at $2 and $10. In independent tests by Artificial Analysis, Astra ranked second behind Anthropic's Claude Fable 5.2. OpenAI's system card rates both models Critical for cybersecurity capability under its Preparedness Framework and says default limits on higher-risk security help have been tightened.

    What it means for you

    If you pay for API use, test Sol against Astra on your own tasks before choosing, since the cheaper model may be enough for routine work. Expect some security-related requests to be refused more often, and treat benchmark claims as a starting point rather than proof.

  3. 03

    Google separates Gemini chat history from AI training, but keeps chats for about two months when history is off

    Google says Gemini app users will be able to keep their chat history while opting out of having those chats used to train its generative AI models; until now both depended on a single Keep Activity setting. If the new History setting is off, Google will save future chats for about two months, which Digital Trends notes is longer than the up to 72 hours under the old setting. Google says these chats are used to respond, process feedback and protect its services. The changes roll out globally over the coming weeks.

    What it means for you

    When the update reaches your account, open Gemini's settings and choose History and training separately instead of relying on earlier choices. If anyone uses a personal Gemini account for work, keep confidential client data out of it, since chats are now kept for weeks even with History off, and Digital Trends notes that sending feedback also shares recent chats.

  4. 04

    Microsoft releases MAI Code 2 and MAI Code 2 Flash, its own coding models, in GitHub Copilot

    On November 17, Microsoft AI released MAI Code 2, aimed at everyday enterprise coding such as debugging and multi-file changes, and MAI Code 2 Flash, for high-volume tasks like code explanations and routine fixes. Both are available in GitHub Copilot, and Code 2 Flash is in limited-access preview in Microsoft Foundry. Microsoft says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of its cost, a comparison based on prices that include a 50% promotional discount, and says the models are not tuned for very long-running agent work.

    What it means for you

    If your developers use GitHub Copilot, the new models are worth a trial on routine coding work, where lower cost may matter more than peak capability. Measure the results on your own code, since the comparisons are Microsoft's own and the current prices are promotional.