Draw your architecture, not a diagram for threat modeling.
The drawing is the model. Context, containers and components on a C4 canvas, with zones and trust boundaries. The threats are worked out from it; there is no second diagram to keep in step.
Continuous threat modeling
Works without AI. Hosted or on your own servers, more than one way to run it is planned. Turn AI on when you want it.
Made for software architects and security architects, working on one model in one place.
Why Tabya
Four things Tabya does today, and what is planned next, marked as such.
Draw your architecture, not a diagram for threat modeling.
The drawing is the model. Context, containers and components on a C4 canvas, with zones and trust boundaries. The threats are worked out from it; there is no second diagram to keep in step.
Fewer false alarms by design, and what it can’t tell yet is kept apart.
Tabya asks before it accuses. A plain question settles what a drawing can’t show, and only confirmed links produce threats. What it can’t decide yet goes on a “To check” list, outside the threats and outside every count. False positives are measured on Tabya’s own reference architectures.
Works without AI, and nothing leaves your machine.
Every threat comes from Tabya’s own rules. The same drawing and answers always give the same result. Today it runs on your machine with no telemetry and no outside calls. AI is planned as something you turn on, never something the engine needs.
Change the drawing, and it tells you what changed.
Tabya reacts. Add, change or remove a part and the threats follow: new ones arrive, unrelated ones stay put, and a decision that relied on a fact that changed is marked for review, with the reason. Every edit gets a short summary.
“Unless the stencils are configured correctly, it is very easy to have tons of false positives.”The maker of Tabya
Next: sign-in, roles and shared editing, so a software architect and a security architect work on one model; then attack paths and an OSI lens. Their previews on this page are marked Planned. Daily vulnerability checks come later.
Drawn for a review, by hand, in a tool made for threat modeling. Every threat it lists was true the day it was drawn. The architecture moves on; the diagram doesn’t. Six months later the question is which of its findings still hold, and nobody can say without drawing it again.
Drawn as the architecture, in the shape architects already use. A new part brings its own threats, a removed one takes them away, and a decision that rested on something that changed is marked for review with the reason. The question six months later is answered by the change history, not by a redraw.
How it works
A web shop with an AI assistant, step by step. The software architect draws it; the security architect works its threats on the same drawing. Steps 7 to 9 are planned, shown as previews.
Draw the system the C4 way: people, containers and, where you want them, the components inside, placed in zones and trust boundaries. Each part is a type from the stencil box, AI agents and MCP tool servers included.
Tabya suggests the links between parts, with direction and protocol filled in. Confirm each one, or dismiss it and it won’t come back. Only confirmed links count.
HTTPSsuggestedconfirmedOIDCsuggestedconfirmedHTTPSsuggestedconfirmedSQLsuggestedconfirmedHTTPSsuggestedconfirmedMCPsuggestedconfirmedHTTPSsuggestedconfirmedSQLsuggesteddismissedThen one plain question at a time, on the part it’s about. Answer yes, no or I don’t know: an unknown never raises a threat and never silences one.
Must callers prove who they are first?
Is the data encrypted when stored?
Does a person confirm each action that changes, sends or deletes something before it happens?
One list from Tabya’s own rules, each threat on its part of the drawing, with its severity and every framework label that fits. The AI agent’s threats also carry MITRE ATLAS and OWASP IDs.
Labels: MITRE ATT&CK (T), CAPEC, CWE; on AI parts also MITRE ATLAS (AML), OWASP’s LLM list (LLM) and agentic list (ASI).
Decide on each threat: planned, mitigated, accepted or not applicable, with a note where one is needed. Each decision remembers the facts it relied on.
The shop adds Payments and opens the assistant to customers. Every edit gets a short summary of what’s new, and a decision whose facts changed is marked for review, with the reason.
Assistant, accepted. “Can anyone on the internet reach this?” on Assistant changed from no to yes.
Paths from anything an attacker can reach first to what you mark as most important, along confirmed links only, each step backed by a threat. Ranked, and worked out again when the model changes.
A layer lens on the same drawing: pick a layer and see its parts, links and threats. Choosing a protocol fills in the layers of its link.
Payments is run by its vendor, so its layers 1 to 5 raise no threats of yours.
Sign-in, roles and shared editing come next, so the software architect draws while the security architect works the threats, on one model.
Today one person works on a model at a time.
Who it’s for
The architecture is the model. Whoever draws the system and whoever works the threats look at the same drawing, and when it changes, both see what changed.
The system the C4 way: context, containers and components, with zones and trust boundaries. Tabya suggests the links; you confirm them and answer a plain question or two about each part. No threat framework to learn first.
One list, worked out from that drawing: each threat with its severity, its place on the canvas and its STRIDE, MITRE ATT&CK, CAPEC and CWE labels. Decide on each one, and the decision stays with the model.
There is no separate threat diagram to keep in step. Change a part and the threats follow, a decision that relied on it is marked for review with the reason, and the change summary says what moved, in words both of you read.
The core is built for one person on one model. Sign-in, roles and shared editing are next, and this page will say when they arrive.
What it does
The core is built and being tested, and it needs no AI. Working together, the views, the optional AI and the live data come after, in this order.
Planned: the OSI lens
Choosing a protocol for a link fills in its layers. Pick a layer and Tabya highlights its parts and links and filters the threats. The lower layers of cloud and SaaS parts are run by the vendor, so they raise no threats of yours.
7ApplicationHTTPHTTP
6PresentationTLSTLS
5SessionTLSvendor-run
4TransportTCPvendor-run
3NetworkIPvendor-run
2Data link—vendor-run
1Physical—vendor-run
A preview of a planned view. Tabya’s catalogue decides which layers a protocol fills.
Method and limits
A threat list is only as good as what it was worked out from. This is what Tabya’s rules read, where they stop, and what its numbers mean.
The parts, the zones and trust boundaries, the links you confirmed, and the answers you gave. A link Tabya suggested and you didn’t confirm counts for nothing.
Today Tabya reads no source code, no repository, no cloud account and no running system. A part that isn’t on the canvas raises no threat, and a protection you didn’t draw or declare isn’t credited.
A rule fires only on what is known. An unanswered question never raises a threat and never silences one: the check waits on the “To check” list, outside the threats and outside every count, until you answer it.
Its false-positive figures come from its own reference architectures, drawn for the purpose, some of them scored before the rules had seen them. No other tool was measured on them, so they say how Tabya does on its own test set, not how it compares with anyone else.
Built to be trusted with your architecture
The threat list is worked out by Tabya’s rules alone, so the same drawing and answers give the same list every time. No AI sits in that loop; the AI that is planned is something you turn on.
No telemetry and no calls to outside services. The first stage runs on your own machine, with fonts and icons inside the app.
Today it runs on your own machine. A hosted service and an edition installed in your own network are both planned; which ways are offered, and on what terms, isn’t decided yet.
False positives are counted on Tabya’s own reference architectures, including ones the rules were never tuned on. The test set is the measure, not a comparison with any other tool.
Questions
No. Every threat comes from Tabya’s own rules, and the same drawing and answers always give the same list. AI is planned as something you turn on, for help with the drawing and for a match when a product’s vulnerability data is missing. Left off, nothing is missing from the threat list.
Not today. The first stage runs on your own machine, sends no telemetry and calls no outside service; the fonts and icons are inside the app. A hosted way of running it is planned, and before one is offered this page will say what it means for where your model is kept.
More than one way is planned: as a hosted service, or installed on your own servers. Which ways are offered, and on what terms, isn’t decided yet. Today it runs on one machine, for one person.
It’s in production, built by one person, and there is no date. This page will say when it opens. There is no waiting list, because this page collects nothing, not even an email address.
Yes. AI apps, agents, MCP tool servers, model services and agent memory are parts you draw like any other, and Tabya’s own rules find their threats, such as other people’s text steering what an agent does. Those threats also carry MITRE ATLAS and OWASP IDs. Tabya itself runs no model to do this.
Not yet. The core is one person on one model, in one browser. Planned next: sign-in, roles and shared editing come before attack paths, so that the two of them work on the same drawing, each in their own role.
A tabya is a small fort built ahead of the main line of defence.
Threat modeling does the same: it puts the defence in front of the attack, while the system is still a drawing. The mark is drawn from göz and çengel, the eye and hook motifs of Anatolian kilims.
Status
It’s being built by one person at siftha, from public methods and its own rules. It isn’t open yet, and this page will say when it is.