A daily read on artificial intelligence with an Australian and New Zealand lens — what moved at each layer of the stack, and what it obliges an AUNZ organisation to do about it.
A sibling to the Cyber Digest, not a section of it. Where the cyber daily slices the day by industry sector, this one slices it by layer — because AI news has no natural industry axis. Every item is rated on independent axes, and a theme with no qualifying item is omitted rather than padded.
Each card ends in a single-line readout. Colour on the digest means impact severity and nothing else; every other rating carries a glyph, so a hue never has to mean two things at once.
Low · Guarded · Elevated · Severe · Critical. How far the risk or obligation position moves.
Proposed and Signalled both mean nothing is enforceable yet.
Also Probable, Other, Independently evaluated and Unreplicated preprint.
Rated only where the story moves capability rather than policy, market or law.
Tier 1/4–4/4 carries a short note on what kind of source it is. Hover the segment for the full rationale.
Verified means established by first-party disclosure, a regulator, or two or more independent sources.
A cyber advisory can be independently verified — a register entry, a patch, an exploitation observation. A frontier capability claim usually cannot be verified at all: the strongest available statement is that the vendor said it, or that someone with no stake measured it. So every item carries a second badge recording what kind of claim it is — vendor claim, independently evaluated, unreplicated preprint, corroborated or rumoured — and no vendor benchmark is ever presented as an established fact.
That is also why there is no equivalent of a CISA advisory here. Nothing in AI carries a statutory duty to publish, so a non-disclosure is caught by no anchor at all. The digest is built knowing that. A vendor's claim about its own model is always a vendor claim — price, speed and benchmark numbers are claims, not measurements.
The window is 48 hours (72 across a weekend) and it is stated in the header of every edition. Outlets report an action days after it happens, so where the dateline and the event date differ, both are given and the event date governs. A story that ran in either of the previous two editions is suppressed unless the day genuinely adds a fact, and the new fact must be what the card is about.
A story older than the window is not a full entry. If it still matters it appears in the executive summary as context, never as a card. A checked silence — a register swept and found quiet — is a finding and is stated; an unchecked silence is a gap.
First edition published 8 October 2026. The format, taxonomies, source register and rating framework are settled. What is not yet automated: the collection pipeline and its own database are still to be built, so editions currently depend on an operator run rather than a scheduled job. Two weak spots are known and stated rather than hidden — Impact, Obligation and Evidence are editor-assigned per card, and the frontier-lab status pages need a dedicated collector before they can be trusted to fire without a human reading them.