Introduction
Here's a roundup of AI news from the evening of July 18. The biggest story: Netflix has acquired InterPositive, an AI startup involving Ben Affleck, for $587 million (roughly 85 billion yen) — a symbolic moment for large-scale AI investment in the entertainment industry. We also cover Kimi K3 overtaking Claude Fable 5 to claim first place on a spreadsheet benchmark, the problem of AI spam filters being fooled by classic "text salting," VideoChat3's bid for video-understanding SOTA with just 4B parameters, and a hands-on investigation into whether Claude Code really commits and pushes .env files.
1. Netflix Acquires "InterPositive," a Ben Affleck-Backed AI Startup, for $587 Million
Variety reports that Netflix, the film and streaming giant, has acquired **InterPositive**, an AI startup involving Ben Affleck, for $587 million (roughly 85 billion yen).
What stands out is the sheer scale of the deal. Generative AI applied to video production has been entering practical use over the past year, and Hollywood is accelerating its push to bring AI startups in-house. For Netflix, the acquisition looks like a technological investment aimed not just at making scripting, editing, and VFX workflows more efficient, but at redesigning the production pipeline for streaming content itself.
That said, the post-acquisition integration schedule and the technical details of InterPositive have not been made public at this point. Hurdles unique to the entertainment-meets-AI intersection remain — including the risk of backlash from creators and relations with existing labor unions — so more information is needed before we can judge how this turns out.
2. Kimi K3 Takes First Place on SpreadsheetBench 2, Overtaking Claude Fable 5
The community is abuzz over **Kimi K3**, Moonshot AI's large-scale model, taking first place on AfterQuery's **SpreadsheetBench 2** and outscoring Anthropic's Claude Fable 5 (via the LocalLLaMA forum).
What makes this noteworthy is that it's an evaluation on spreadsheets — a "practical, unglamorous" task. LLM benchmarks tend to skew toward reasoning and coding, but spreadsheets serve as a meaningful proxy for real-world business utility. Kimi K3 has also ranked highly on agentic benchmarks such as DeepSWE, which suggests consistent strength on tasks closer to real work.
That said, for the evaluation design and sample size behind SpreadsheetBench 2 you'll need to consult AfterQuery's published information, and it's worth avoiding firm conclusions about which model is better based on rank changes in a single benchmark alone.
3. AI Spam Filters Get Bypassed by "Text Salting"
The Register reports that generative-AI-based spam filters are easily bypassed by the classic technique known as **text salting**.
Text salting is an old trick: mixing random words or unrelated passages into the body of a spam email. It was originally used against rule-based filters, but it has now been confirmed to work against LLM-based filters too. The likely explanation is that AI filters weight "contextual meaning" so heavily that they end up being pulled along by cleverly disguised noise.
The structure here — a new layer of defenses reactivating an old layer of attacks — is a pattern seen over and over in security. If you're putting AI-based decision systems into production, it seems regression testing against legacy attacks needs to be redesigned from scratch.
4. VideoChat3: A Fully Open MLLM Aiming at Video Understanding with 4B Parameters
A paper on **VideoChat3**, a multimodal LLM specialized for video understanding, has shown up on Papers with Code.
Its defining feature is the design: despite having only **4B parameters**, it aims for performance on par with much larger models. The authors improve efficiency with two mechanisms — I3D-ViT (Inflated 3D Vision Transformer) and Adaptive Frame Resolution — while newly constructing three training datasets: VideoChat3-Academic2M / LV116K / OL617K. This strengthens generalization across three scenarios: general video, long-form video, and streaming video.
The release is billed as "fully open," with training code, training strategies, and the datasets all expected to be provided. The frontier of video understanding has so far been led by big, closed models, which is exactly why a small, reproducible open implementation could be of significant value to the research community.
5. Does Claude Code Really Push .env Files to Public Repos? — A Hands-On Test on Qiita
A Qiita article is drawing attention: its author actually tested, in their own environment, whether Claude Code (Anthropic's CLI coding tool) will mistakenly push `.env` and other files containing API keys to a public repository when asked to "commit and push."
In our previous digest, on the morning of July 18, we covered the fix for a permission-check vulnerability in Claude Code v2.1.214. This time the author set out to confirm "how safe it is in real operations when permissions are configured correctly" — so it's positioned not as a story about the vulnerability itself, but as an empirical report on the operational side.
How much handling of secrets can be delegated to AI agents is a perennial concern for every engineer designing automation pipelines. For the details of the measured results, be sure to read the original article.
Wrap-Up
That's it for the afternoon-to-evening edition. The lineup skewed toward topics close to practical work: a major M&A deal (Netflix × InterPositive), the fine print of model evaluation (Kimi K3), an AI security blind spot (spam filters), open research (VideoChat3), and an operational verification (Claude Code). There may be more movement late tonight, but that's a wrap for today's daytime-to-evening edition.