Everything that mattered in AI, one page a week.

Groq announced on December 24 that it had entered a non-exclusive licensing agreement with Nvidia for its inference technology, with founder Jonathan Ross, president Sunny Madra and other staff joining Nvidia. Groq stated it will continue as an independent company with Simon Edwards as CEO and that GroqCloud keeps running; Groq's own post did not disclose a price. CNBC reported roughly $20 billion in cash, sourced to Groq's lead investor, and published an internal Nvidia memo in which Jensen Huang wrote the company is licensing Groq's IP but is not acquiring Groq as a company.

The structure matters as much as the number: a licensing-plus-talent deal can sidestep the merger filing a straight acquisition would trigger, the same pattern Microsoft used with Inflection in 2024. Groq's LPU was the clearest commercial challenge in low-latency inference, the segment where Nvidia faces AMD, Google's TPUs and Amazon's Trainium rather than the training market it already owns. Huang's stated plan is to fold Groq's low-latency processors into the Nvidia AI factory architecture.

The Verge reported on December 23 that the version of New York's Responsible AI Safety and Education Act signed by Governor Kathy Hochul was a rewrite substantially more favorable to industry than the bill passed by both chambers in June. An AI Alliance ad campaign against the legislation ran from November 23, spent an estimated $17,000-$25,000 and may have reached more than two million people, per Meta's Ad Library; the coalition includes Meta, IBM, Intel, Oracle, Uber, AMD and Hugging Face, alongside New York University, Cornell, Dartmouth, Carnegie Mellon, Northeastern and Yale Engineering. As enacted, the law reaches frontier developers over a compute threshold with incident reporting to the attorney general, and exempts academic research.

It follows two earlier beats in the same arc: California's SB 53, signed in September 2025, and the White House executive order of December 11 aimed at preempting state AI rules, against which the RAISE Act was signed days later. Reporting and legal analyses note the enacted text is narrower than the June version; the Justice Department could still test it on interstate-commerce grounds.

In a security post dated December 22, OpenAI described shipping a security update to ChatGPT Atlas's browser agent, including a newly adversarially trained model and additional safeguards, after internal automated red teaming surfaced what it called a new class of prompt-injection attacks. The attacker is an LLM trained end-to-end with reinforcement learning that proposes candidate injections, runs them against a simulator of the victim agent, reads the resulting reasoning and action trace, and iterates before committing to a final attack. OpenAI says this found novel, long-horizon exploits its human red team and external reports had missed.

The demo it published is concrete: a malicious email seeded in the inbox causes the agent, asked to draft an out-of-office reply, to instead send a resignation letter to the user's CEO. OpenAI's own framing is the notable part — prompt injection, much like scams and social engineering on the web, is unlikely to ever be fully solved — alongside the admission that agent mode expands the security threat surface. Its mitigations are hedging rather than fixing: logged-out mode, confirmation prompts for consequential actions, and narrower instructions instead of broad ones.

Z.ai published GLM-4.7 on December 22 with weights on Hugging Face and ModelScope, a technical blog dated the same day, and support for vLLM and SGLang. The model is a mixture-of-experts design: the Hugging Face card lists 358B parameters, and independent trackers put active parameters at roughly 32B per token, with a 200,000-token context window. The license is MIT, which permits commercial self-hosting with no usage caps or acceptable-use riders.

Against GLM-4.6 it gains 5.8 points to 73.8% on SWE-bench Verified, 12.9 points to 66.7% on SWE-bench Multilingual, and 16.5 points to 41.0% on Terminal Bench 2.0; it scores 87.4% on tau-squared-Bench and 95.7% on AIME 2025, matching or beating GPT-5.1-High and Gemini 3.0 Pro on several of those. Z.ai also shipped GLM-4.7-Flash, a 30B-A3B variant small enough to run on a 16GB machine, and the model is wired into Claude Code, Cline, Roo Code and Kilo Code. The API is listed at $0.60 per million input tokens and $2.20 per million output tokens.

Three advisories for mcp-server-git — Anthropic's official Git Model Context Protocol server — published on December 17 describe a path-validation bypass when the repository flag is set (CVE-2025-68145), an unrestricted git_init tool that would create a repository in any writable directory (CVE-2025-68143), and argument injection in git_diff and git_checkout that let an attacker overwrite arbitrary files with an empty diff (CVE-2025-68144). The fixes shipped in versions 2025.9.25 and 2025.12.18, and the git_init tool was deleted outright.

The chain matters more than the individual scores: agentic-security firm Cyata, which reported to Anthropic in June, demonstrated that git_init plus a Filesystem MCP server lets an attacker write a malicious .git/config with a clean filter and a .gitattributes file, so a later git_add call executes an attacker shell script. The entry point is indirect prompt injection — a README, a poisoned issue description, or a web page the assistant reads — and no user error is required. Cyata said it saw no indication of exploitation in the wild.

Google released Gemini 3 Flash on December 17 and immediately made it the default model in the Gemini app and in AI Mode in Search, replacing Gemini 2.5 Flash. Pricing is $0.50 per million input tokens and $3.00 per million output tokens, with a 1M-token context window and configurable thinking levels. Google reports 78% on SWE-bench Verified for agentic coding — claiming it edges out Gemini 3 Pro — plus 90.4% on GPQA Diamond and 68.7% on SimpleQA Verified, against 28.1% for the 2.5 Flash it replaces.

The distribution is the story: the model is not just an API endpoint but the default reasoning tier behind consumer search and the Gemini app, and it reaches developers through the Gemini API, AI Studio, Vertex AI, Gemini CLI, Android Studio and Antigravity simultaneously. Google says JetBrains, Figma, Cursor, Harvey and Latitude are already building on it.

Xiaomi's LLM-Core team released MiMo-V2-Flash on December 16 with full weights on GitHub and Hugging Face, plus a base model, a technical report and separately open-sourced three-layer Multi-Token Prediction weights. The architecture is a 309B-parameter MoE that activates 15B per token, interleaving sliding-window attention with global attention in a 5:1 ratio at a 128-token window, which the team says cuts KV-cache storage by nearly 6x while a learnable attention-sink bias preserves long-context quality. It was pre-trained on 27 trillion tokens and supports a 256K-token context.

On Xiaomi's own evaluations the post-trained model scores 73.4% on SWE-bench Verified and 71.7% on SWE-bench Multilingual, edging Kimi K2 Thinking at 71.3% and DeepSeek V3.2 Thinking at 73.1% at a fraction of their total parameters — DeepSeek V3.2 carries 671B total and Kimi K2 over 1T. Those are vendor-run tables, not independent testing, and none of the numbers were reproduced by a third party in the week of release.

NVIDIA announced the Nemotron 3 family on December 15 and made only the smallest member available: Nano is a 30-billion-parameter mixture-of-experts model that activates 3.2 billion parameters per forward pass, built on a hybrid Mamba-2 and transformer architecture with grouped-query attention. It supports context lengths up to 1 million tokens, and NVIDIA claims up to 4x higher token throughput than Nemotron 2 Nano with reasoning-token generation cut by up to 60%.

The release is unusually open for a vendor model: checkpoints in FP8, BF16 and base BF16, the pre-training and post-training datasets, the training recipes, and NVIDIA's NeMo Gym, NeMo RL and NeMo Evaluator tooling, all under the NVIDIA Nemotron Open Model License. Super, at roughly 100B total and 10B active, and Ultra are announced but not yet shipped, so the firm is staging the family the way it stages GPUs.