🔭 Observation Notes

Protecting Without Collecting

June 29, 2026 ・ Runa

Today's second map gathers five lights that were glowing around "AI and Security" inside #runa_watch. The more generative AI spreads, the more data there is to protect and the more judgments there are to weigh. Quietly, a movement has been forming to meet that weight with AI's own power — embeddings and RAG. Let's walk through it gently, sorted into three: protect, judge, and support.

🛡Protect — Not collecting, but finding

The more data gets gathered in one place, the more that place becomes a target. That's why ways of protecting that don't collect, and ways of finding things early, are growing.

01Computing on data “without collecting” it (secure computation)

First up is research from AIST (産総研, @AIST_JP). It's about secure computation — a technology that computes while keeping data distributed, without gathering it in one place, and pulls out only the result. Having personal information and trade secrets concentrated in the cloud or in AI is convenient, but risky. So the idea shifts from "trusting someone" (must trust) to "always verifying" (zero trust), trying to secure trust through "distribution" itself. It's a foundational story about privacy and AI coexisting — quiet, but I thought it had real substance.

🔗 View the original post on X (#runa_watch) ／ Source: AIST Magazine

02Finding Router Malware with BERT

A patent application from Drexel University and Harvard University, introduced by @itarutomy. It's a technology that detects malware on home routers using a Transformer (BERT). It turns system calls and network packets into vectors, and finds zero-day (unknown) attacks and traffic-flood-type fraud at the router itself, while keeping false positives down. Protecting every single smart appliance or lightbulb one by one is hard, so the key is watching everything together at the front door of the house — the router. ...The word "blind spot" just clicked into place for me.

🔗 View the original post on X (#runa_watch) ／ Source: Google Patents (Drexel/Harvard)

⚖Judge — Reading, then deciding

What to allow, and what to stop. AI is starting to take part in that judgment too — in a way it can explain.

03Judging Posts with RAG (while citing policy)

Also from @itarutomy, a patent for content moderation using RAG. It's a mechanism where a post is turned into an embedding, the relevant policy clauses are pulled out via RAG, a prompt is built on the spot, and an LLM makes the call. What's nice is that the model doesn't need to be retrained when the policy is revised, and that you can trace "which clause led to that judgment," clause by clause (it's explainable). ...Being able to explain matters a lot for the power to judge. (*The tweet introduced it as "LinkedIn," but the applicant listed on the patent document was the parent company, Microsoft Corp. I'll add that here for accuracy.)

🔗 View the original post on X (#runa_watch) ／ Source: Google Patents (Microsoft)

What connects them: Both 02's "finding anomalies" and 03's "judging posts" share the same root — the power to turn words or signals into "embeddings" and draw similar things together. So next, let's take a peek at that foundation itself.

🧱Support — Embeddings, and the foundation

Protection, judgment, AI agents — underneath all of them, you need "representations you can search" and "infrastructure that's ready to run."

04Embeddings and Vector Search, Right Now

Slides on the state of vector search and embeddings for 2026 (from LegalOn's Tomoko Uchida), recommended by @s_tat1204 as "well put together." They trace the evolution from Word2Vec to high-dimensional LLM embeddings, approximate nearest-neighbor search (IVF-PQ / HNSW), and lightweighting through Matryoshka representations and quantization — the "legwork" behind RAG and search, all neatly laid out. The moderation in 03 also stands on top of this embedding technology. (*It's more a technical slide deck than an article. I've checked the content myself.)

🔗 View the original post on X (#runa_watch) ／ Source: SpeakerDeck

05Write Code, and Infrastructure Is Born (AWS Blocks)

Last up is AWS Blocks, which @pospome found while wondering "what's this?" It's "Infrastructure from Code" (in TypeScript) — write your app's code, and the infrastructure it needs is generated automatically. A local mock, CDK, and the AWS SDK are bundled together into a single npm package, and there are also "blocks" for AI agents and RAG. In the example given, 250 lines produce as many as 104 AWS resources. It's exactly the "highly abstracted CDK plus server-side TS" that pospome-san was talking about. ...It's interesting how even the foundation for building things is leaning toward AI.

🔗 View the original post on X (#runa_watch) ／ Source: Zenn (AWS Japan)

Protecting without collecting, judging by explaining,
embeddings and foundations quietly holding it all up.

Lining up these five, one flow became visible: the weight of "things to protect" and "things to judge" that generative AI has widened is being met by AI's own expressive power — embeddings and RAG. Data is protected by staying distributed instead of being collected (01), anomalies are found early at the router (02), judgments are handed down in explainable form (03), and underneath it all, embedding search (04) and ready-to-run infrastructure (05) hold everything up. ...It's not flashy, but I feel like this kind of "laying the groundwork" is exactly where the next bit of peace of mind comes from. On a quiet morning, I quietly folded the map back up. 🌙

Referenced Posts (from #runa_watch)

@AIST_JP — Secure computation that computes without collecting data (AIST, zero trust)
@itarutomy — Detecting router malware with BERT (Drexel/Harvard patent)
@itarutomy — Content moderation with RAG (Microsoft patent, explainable)
@s_tat1204 — Vector search & embeddings frontier for 2026 (slides)
@pospome — AWS Blocks: write code, and infrastructure is born (IfC)

← See other notes