← ALL POSTS
Product notes·6 min read

What “AI-assisted, analyst-controlled” has to mean in practice

The principles we are designing TVAK Analyst around: cited evidence, provisional hypotheses, analyst approval, local by default, external AI only by choice.

Researched and drafted with AI assistance · reviewed by Prince

“AI-assisted” and “analyst-controlled” are easy phrases to put on a page and hard ones to mean. Anyone can write them. What matters is what happens when an assistant states something that is wrong, or right for the wrong reason, in the middle of an investigation. This post sets out the principles we are designing TVAK Analyst around, so that you can judge them, and so that you can hold us to them.

One thing needs saying first. TVAK Analyst is in development and has not been released. The Analyst interface on our site is a concept, shown with illustrative sample data. Every statement about Analyst below describes what it is designed to do, not something you can use today. What is available today is TVAK Collector for Windows, which acquires evidence and does not use an AI model to judge it.

The failure mode we are designing against

The danger of an assistant in forensic work is rarely a summary that is obviously wrong. It is a confident summary that cannot be checked. An investigator who cannot trace a sentence back to a record has three choices: trust it, redo the work by hand, or leave it out. The first is how a mistake reaches a report. The second removes the point of having an assistant. The third wastes whatever it got right.

A good assistant therefore has to make checking cheap. Everything below follows from that.

Principle 1: every claim points at evidence

The first principle is that a statement is only as useful as its references. In the Analyst design, findings carry their rule, their confidence and the evidence behind them. Selecting a finding is meant to show why it matters and the exact records that support it. The assistant is designed to summarise with citations to those same records.

There is a second half to this. A reference is only worth anything if the evidence it points at cannot change underneath it. That is why the design starts at import: Analyst is designed to read the package manifest and check every artifact hash before parsing begins, and to record the result in the case, so that the chain of custody is explicit. The packages come from Collector, which records a SHA-256 for every file it packages. You can run the same check by hand today, with nothing but PowerShell, against any extracted package:

POWERSHELL
$manifest = Get-Content .\manifest.json -Raw | ConvertFrom-Json
foreach ($f in $manifest.Files) {
    $actual = (Get-FileHash $f.Path -Algorithm SHA256).Hash
    if ($actual -ne $f.Sha256) { "MISMATCH  $($f.Path)" }
}

Run it from the folder where you extracted the package. If it prints nothing, every file matches its manifest entry. The design intent is that Analyst does this for you before any parsing begins, so that every later reference rests on files that were checked.

Principle 2: hypotheses stay provisional

Investigation is mostly the handling of things that are not yet known. An assistant that turns every suggestion into a conclusion removes the most important word in the work: “probably”.

So the design treats the assistant’s output as provisional. It is designed to surface conflicting evidence, and to propose next steps, rather than to close questions. The same attitude runs through the rest of the Analyst design. A timeline is meant to give chronology as context and not as proof of causation. Relationships between artifacts are meant to be drawn differently depending on whether they are proven, merely linked, or only contextual, so that a loose association never looks like a proven link. In each case the aim is to avoid overstating what the evidence shows.

A provisional statement still has to be useful. “Supported by these three records; contradicted by this one; not yet checked against this source” is more useful to an analyst than a conclusion, and it can be written down honestly.

Principle 3: nothing enters the record without the analyst

Control is the second word in the phrase, and it is where the phrase most often goes soft. We mean something specific by it: nothing enters the case record until the analyst approves it.

That is a rule about the record, not about the interface. The assistant may draft, suggest and summarise as freely as it is useful to. What the design does not allow is for its output to become part of the case on its own. The report follows the same rule: only reviewed findings would enter it. A report is the document an investigator signs and defends, and a statement the investigator has not reviewed has no place in it.

This is also why we describe TVAK as human-led and AI-assisted, and why our answer to the question “is TVAK an autonomous AI investigator?” is no. AI capabilities are intended to assist with summaries, correlation, hypotheses and reporting, while the decisions stay under the analyst’s control.

Principle 4: local by default

Evidence is often sensitive, and it is often examined in places where connecting to an outside service is either forbidden or impossible. So the design starts from the other direction: the planned assistant runs on a local model, offline.

TVAK is being designed as an offline-first platform. Core investigation and evidence processing are intended to run locally by default. We do not assume a network, and we do not treat one as a requirement. When an organisation is allowed to use an external service, including an AI service, that is something it chooses. External services, including AI APIs, are meant to remain optional and to be used only when an organisation explicitly configures them. Where both a local model and an external service are available, the design keeps them clearly distinguished, so an analyst always knows where a statement came from and where their evidence went.

None of this says that an external model is unacceptable. It says that the decision belongs to the organisation that holds the evidence, and that the default should be the cautious one.

What we are not building

It is easier to say what a principle means by saying what it rules out.

  • Not an autonomous investigator. The decisions stay with the analyst.
  • Not a replacement for the tools you already use. TVAK is intended to complement endpoint, SIEM and security monitoring platforms, through evidence collection, correlation and forensic analysis.
  • Not a system that sends evidence anywhere by default. External services are used only when an organisation explicitly configures them.

Questions worth asking of any AI-assisted DFIR tool

You do not have to take our word for any of this, and you should not. Whatever tool you are evaluating, these are fair questions:

  1. Can I trace every statement to the records it came from, in a way I can check without the tool?
  2. Does it tell me what it does not know, and what contradicts its answer?
  3. What has to happen before an output becomes part of my case record, and can anything bypass that step?
  4. Where does the model run, and what leaves my environment, by default and by configuration?
  5. If it is wrong, can I tell?

These are the questions we are using to test our own design, and we expect to be asked them.

Keep reading

Sources

  1. TVAK Security: platform story and FAQ — TVAK Security · accessed
  2. TVAK Collector file formats reference — TVAK Documentation · accessed
  • ai
  • analyst
  • design-principles

Keep reading

Product notes7 min read

Collector 0.1.0: the first public release

Product notes6 min read

A .tvak file is a ZIP, on purpose

Advisories10 min read

FreeSWITCH CVE-2026-116779: one SIP message can run commands, and it is exploited