← Back to the news

How this is built

What is materially new in the privacy, security, legal and risk dimensions of AI: agentic systems, generative AI and LLMs, models and the model supply chain, and robotics and embodied systems?

Register last reviewed 2026-08-13 · 56 sources

How sources are weighted

Every source sits in one of three tiers. A claim rests on primary sources; the rest widen what gets seen and help confirm a story.

Primary

Original reporting or a primary document: research papers, CVEs and vendor security advisories, regulator and standards-body texts, first-party vendor engineering/security posts, and incident post-mortems. Weighted highest.

Secondary

Journalism and analysis that reports on primary sources: the security trade press and analyst blogs. Used for discovery and corroboration, not as the sole basis for a claim.

Tertiary

Aggregators and community surfaces (subreddits, link aggregators, roundups). Used to widen recall and surface items the curated feeds miss; every hit is confirmed against a primary or secondary source before it enters the store.

What gets in

What stays out

How a story gets here

The counts here are read out of the running configuration when this page is generated. They describe the pipeline as it stands today.

  1. Collect. A scheduled job runs once a day, reads all 54 feeds and 2 query APIs, and keeps anything published in the last 10 days.
  2. Filter on keywords. A plain text match, with no model involved. 38 feeds must match both an agentic-AI term and a privacy, security or legal risk term. 13 regulator and standards feeds need only the risk term, because their wording rarely says "agent". 3 low-volume release feeds skip the gate.
  3. Drop what has been seen. URLs are normalized and compared against everything already stored. A story that circulates for a week is recorded once.
  4. Select and summarize. What survives goes to claude-opus-5, which picks what to keep and writes the summary and the per-lens notes. Any item whose URL is not taken verbatim from the candidate list is discarded before it is stored. The model cannot introduce a source of its own.
  5. Score on three lenses. Each item is rated 0–3 for privacy engineering, security, law and risk independently, given an overall importance of 1–5, and tagged from fixed vocabularies of 11 security, 11 privacy, 10 legal and 10 risk subtopics. Tags outside those vocabularies are discarded.
  6. Write the digests. A daily digest is generated for each lens. Once a week a deeper pass fetches the articles and papers themselves and reads them in full.

Each card carries a full text read or from abstract marker. Day to day, summaries are written from the title and abstract the source published in its own feed. Only the weekly pass opens the article itself.

What this cannot do

The keyword gate in step two is the weakest part. An attack described in language nobody uses yet will not match, and nothing downstream can recover a story that was never collected. Reviewing the register on a cadence is partly meant to catch that.

The summaries are written by a language model. They can be wrong, or can flatten a detail that mattered. Every card links to its source, which is the authority. Coverage is English-language and weighted toward the sources below, mostly EU and US regulators.

The sources

The full register, kept in the open. Each daily run reads these feeds plus the query APIs below, then an AI pass selects and summarizes what is genuinely new. Most feeds are keyword-filtered first for agentic-AI privacy, security, and legal relevance. Primary regulator and standards feeds skip that gate and go straight to the AI pass, since their wording rarely says "agent".

Primary 29

Secondary 23

Tertiary 2

Query APIs 2

Keeping it honest

The register is audited on a regular cadence. The audit checks that every feed is still reachable and flags the ones that have gone quiet. It also looks for gaps the current list would miss. Feeds get added, moved between tiers, or dropped on that evidence, and the change takes effect on the next daily run.

The register is a single configuration file, and this page is generated from it on every run. The source list, the counts, the filter split and the model name above are read out of the live configuration. Add a feed or change a rule, and this page changes on the next build.

See also the privacy page.