How it works
Many sources. One score. No black box.
Defendr checks one item against dozens of independent security sources, combines what they say with fixed weights, and then explains the result. AI is used to add public context around an item — never to invent the verdict.
The pipeline
Five stages, start to finish.
A scan is a fan-out and a merge: ask many sources at once, weight what each one says, then show the working — with AI doing the open-web searching and the plain-language phrasing.
- 01
Detect what you pasted
One box accepts a link, a domain, an IP address, a file hash or a program name. The input is normalised — http vs https, trailing paths, lookalike domains — and the type is decided by fixed rules rather than AI, so the lookup can't be steered into the wrong database.
- 02
Fan out to every source
A single check queries VirusTotal's 70+ engines, the abuse.ch malware repositories, the OpenPhish feed, a filtering DNS resolver, the domain registry and live DNS records — all at once, in parallel. AI covers what those databases can't: it searches the open web, opening reviews, forum threads, app-store listings, news and security write-ups, and extracts only what bears on trust.
- 03
Weigh the evidence
Each signal carries a fixed, published weight. Engine disagreements, repository listings and community reports are combined into one 0–100 score with plain arithmetic — deliberately AI-free, so a confident-sounding model can never move a verdict on its own.
- 04
Explain it in plain language
The report names every engine that flagged the item and shows the blocklists that listed it, while AI phrases the research in everyday words: a two-to-four-sentence summary, a sentiment of trusted, mixed, risky or unknown, up to four positives and four red flags, and the sites it relied on.
- 05
Layer in human context
Community reports feed back into the score, weighted by what kind of report it was, with reporter badges alongside them. No AI is involved here — this is people reporting what happened to them, kept separate from machine output so you can always tell which is which.
Sources
Every source, and what it is allowed to contribute.
Some sources can only add context. Others can move the score, and they do so by a fixed amount — so a single noisy feed cannot sink an otherwise clean item.
The score
How the number is actually calculated.
The score is plain arithmetic over the sources above, clamped to 0–100. Reproduce it by hand if you want to — that is the point.
The share of engines that called it unsafe, scaled to a 52-point ceiling.
3 extra points per engine that returned a named malware detection, capped at 20.
A fixed 26–34 points per listing, depending on the source's reliability.
+6 per abuse report (max 15), −4 per “this was fine” report (max −10).
Nothing here is averaged away or hidden: the total is clamped to 0–100 and every point can be traced back to the source that contributed it.
Reading the number
Unknown means none of the sources had anything to say — not that the item is safe. A brand-new link, for instance, has no history to judge.
The AI layer
AI reads the internet so you don’t have to — and nothing more.
Technical sources answer “is this known bad?”. They cannot tell you whether a download is adware bundled into an installer, or whether a shop takes your money and never ships. That is what the optional reputation research does: it searches the public web for what people and researchers actually wrote about the item, then condenses it.
Searching the public web for context
For a program, domain or link, an AI research pass reads reviews, forums such as Reddit and Stack Exchange, app stores, news coverage and security write-ups, then answers three questions: what is it, who runs it, and has anyone documented abuse?
Turning scattered opinions into a structured answer
The pass returns a fixed shape every time — a short plain-language summary, a sentiment of trusted / mixed / risky / unknown, the publisher, up to four positives, up to four red flags and the sites it relied on. No free-form rambling you have to interpret.
Staying honest about confidence
It is instructed to state only what the sources actually support, to name the sites it used, and to answer “unknown” when a name is generic or the results describe several different things. That is why the panel can say “I couldn't tell” instead of inventing a verdict.
What AI does here
- Reads the public web
- Summarises in plain language
- Names its sources
- Says “unknown” when unsure
What it never does
- Guess a verdict
- Touch the risk score
- Invent a source
- Sit in the dark about uncertainty
- It never changes the risk score — the number stays deterministic and auditable.
- It does not look at the engine data; VirusTotal's verdicts are reported as they are.
- Your uploaded file's contents are never sent to the AI — only a name, domain or link.
- It runs only when you ask for it, and only for signed-in users, because it spends a lookup.
- Results are cached for a week, so the same question is not paid for twice.
- Conflicting findings are surfaced as red flags, not smoothed over.
Read the summary as a starting point, not a guarantee. AI-assisted research can miss things or repeat an opinion it found; the engine data and blocklists are the hard evidence, and the community panel shows what real people reported.
Under the hood
Designed to be quick, cheap and private.
Your file stays on your device
Choosing a file fingerprints it locally in your browser. Nothing is uploaded unless you explicitly ask for a full analysis, and uploaded files are stored privately.
Answers come from cache when they can
Clean results are reused for 12 hours, anything else for 4 hours, and in-progress analyses for a few minutes — fast reports that stay inside API rate limits.
Rate limits are respected, not hidden
VirusTotal's free tier allows a handful of lookups per minute. When the budget runs out, you get a clear message instead of an empty report.
Keys stay on the server
Every lookup runs on the server, so no API key or credit is ever exposed to the browser.
Still wondering about something specific?
Scan it. The report shows each source, its verdict and the exact points it contributed.