Skip to content
IIAIT

AI Today · 13 November 2026

Google releases Gemini 4 Hydrogen, and OpenAI launches Daybreak tools that find and patch security flaws

Google ships a faster, lower-cost Gemini model across its API and apps, OpenAI adds two security products built on agents, Anthropic lets companies run Claude Managed Agents on their own infrastructure, and IBM previews small models trained to resist instructions hidden in documents.

  1. 01

    Google releases Gemini 4 Hydrogen, a faster and cheaper model for coding and agent tasks

    Google introduced Gemini 4 Hydrogen on November 12, 2026, describing it as its everyday model for coding and agentic work, with Gemini 4 Carbon and Argon kept for deeper reasoning. It is available in the Gemini API and Google AI Studio, Android Studio, Gemini Enterprise, Google Search and the Gemini app, where free users get it through automatic model selection. API pricing is $1.50 per million input tokens and $7.50 per million output tokens across a context window of 1 million tokens. Google says it runs about 2.5 times faster than Gemini 3.8 Flash and makes more than five times fewer fabricated claims; these figures come from Google.

    What it means for you

    If you pay for a larger model for routine drafting, coding or document tasks, compare Hydrogen on a sample of your own work, since the price gap can be significant at volume. Treat the accuracy claims as the vendor's own and keep checking facts in outputs that go to clients.

  2. 02

    OpenAI launches Daybreak Red Team and Defense Factory to find and fix software vulnerabilities with agents

    OpenAI announced two products on November 12, 2026 under its Daybreak security platform. Daybreak Red Team, in early access, uses groups of agents with source code and application context to probe web applications for weaknesses, check vulnerability reports and test the resulting patches. Daybreak Defense Factory scans codebases with agents running on the GPT-6 Cyber model and drafts fixes as pull requests with a test, which engineers review and approve before release. According to The Register, OpenAI did not give prices for either product.

    What it means for you

    Automated testing and patching can shorten the time between finding and fixing a flaw, but the review step matters: keep a qualified person approving each change and confirm any scanning is authorized and limited to systems you own. If you buy security services, ask vendors how they validate findings produced by AI.

  3. 03

    Anthropic opens a private beta for self-hosted Claude Managed Agents

    Anthropic said on November 12, 2026 that companies can now run Claude Managed Agents in their own cloud or on-premises environment, in a private beta requested through their Anthropic account team. The customer hosts the agent harness, sandboxes, credentials and session history, and chooses whether model requests go directly to Anthropic, through Amazon Bedrock, Google's Agent Platform or Microsoft Foundry, or through its own gateway. Credentials are kept outside the sandbox, each environment can be limited to an allowed list of hosts, and every agent step can be exported to existing monitoring tools. As Help Net Security notes, the Claude model itself still runs outside the customer's network.

    What it means for you

    For firms with strict data rules, such as in finance or law, this can make it easier to keep agent session records and credentials under their own control. It also moves the operating work to your team, so weigh the hosting effort against your actual compliance needs before asking for access.

  4. 04

    IBM previews Granite models trained to separate instructions from data to resist prompt injection

    IBM published two experimental Granite 4.1 models on November 12, 2026 that use LLMON, a markup format that tags instructions and data separately so the model only follows instructions marked as such. On the OPI benchmark, IBM reports that the preview models resisted nearly all instructions hidden in documents, avoiding 47% more attacks than the best frontier model it tested. IBM says the resistance is learned rather than guaranteed, that adaptive multi-turn attacks are still being studied, and that the models are for evaluation, not production.

    What it means for you

    Prompt injection, where a document or web page contains hidden instructions for the AI, remains a main risk when assistants read outside content. Until such defenses mature, limit what an assistant can do after reading untrusted files, such as sending email or changing records, without your confirmation.