A link's intent, judged in milliseconds, before the message is delivered.
MISUIC is a URI intent classifier that sits inside the mail path. It does not just check a link against a list of known-bad addresses: it asks what the link is actually for, who it claims to represent, and whether its own domain has earned the right to make that claim, then answers before the message is delivered, with no model call and no GPU.
Why it exists
On 2026-08-05, the mail pipeline's own AI content classifier looked at a
link to www.westpac-account.com and rated it "safe, a
legitimate secure login page for Westpac." Its reading of the page was
accurate: the extraction correctly identified the company name and the
purpose of the page. Only the judgement was wrong, because that domain has
never belonged to Westpac.
MISUIC exists to take that judgement away from the model. A domain either belongs to the brand it claims, or it does not, and that question does not need an opinion: it needs a deterministic check, run the same way every time, in front of the model rather than instead of it.
What it decides
MISUIC returns one rating per message, the worst rating carried by any URL it contains, so a single condemned link outweighs nine clean ones:
The URL impersonates a brand it does not belong to, or its host is already confirmed bad from earlier sightings.
Nothing has earned trust and nothing has condemned it either. This is the default: safe is never assumed, only earned. Today, unknown is where an inconclusive URL rests; a further cross-check for exactly these cases is planned but not yet built.
Every URL in the message resolves to a domain with positive standing, a known brand domain or a confirmed-good host.
The message had no URLs at all, so nothing is added to it.
See how it works for the full rule ladder behind these ratings, and how "fail" (our problem, never the message's) keeps the mail path open when MISUIC itself has trouble.
Backed by SURIBS
MISUIC's own judgement is structural: it asks whether a domain belongs to the brand it claims. It is corroborated by SURIBS, the hostname and domain reputation service published from the same ecosystem, which supplies the history: a host already confirmed bad from earlier sightings, or a host quietly earning trust over time.
Structure, then history
MISUIC's rules run first and can condemn a link no reputation source has ever seen. SURIBS then either corroborates that finding or lets a genuinely unproven domain stay at "unknown" rather than being condemned on structure alone.
One cache, two consumers
The same confirmed-bad and confirmed-good entries MISUIC reads are also read by the wider content-classification pipeline it sits in front of, so a host confirmed once is never re-litigated by every consumer that later sees it.
Design principles
- Deterministic before probabilistic. A model's reading of a page can be accurate and its judgement still wrong. MISUIC's checks run first, and structural evidence outranks a model's opinion.
- Safe is earned, never assumed. The absence of bad evidence is "unknown," not "safe." Only a domain with positive standing, a known brand domain or a confirmed-good host, reaches "safe."
- Fail open, always. Any internal error is MISUIC's problem, never the message's: mail keeps flowing exactly as if MISUIC were not there.
- CPU first, GPU added deliberately. Every rule MISUIC runs today is deterministic: no model call, no contention with the content classifier for a shared GPU slot, an answer inside the mail path in single-digit milliseconds. This follows the same principle as the rest of Michelle's mail-filtering work: run on modest, older hardware first, and add newer hardware and GPU-backed checks deliberately, for accuracy and coverage, only where they earn their keep. A brand-check verdict on its own is a first-pass signal, not a final word: a GPU-backed cross-check for the URLs that stay inconclusive after these rules is planned as that next step, not live yet.