Ask about a capture, and check every answer
A model on your computer or on your own cloud account reads what Frame Bench has decoded, and every frame, stream or finding it cites is looked up in the capture. You read the exact text before anything is sent.
Included in the Complete bundle, in the desktop app. The web edition has no AI features, and the assistant is off until you set a model up.
EnlargeHow a question goes
Four steps, and the second one is yours: nothing is sent before you have read it.
Ask
Open the panel from the Analyze menu, or from the menu of a frame, stream, finding, Attack story node or file.
Type a question, press Triage this capture, or click one of the starter questions. Type / to ask for a tool by name.
Read what will be sentThe one step that sends
A preview lists the destination, the size in tokens, an estimated cost and what was replaced.
Send is the only way anything goes. A cloud service always shows the preview for the first question of a conversation.
The tools run on your computer
The model asks for tools; they read the decoded capture here, and each result is redacted before it goes back.
Each step is shown as it ends, with its time. The tools that read the whole capture stop after about seven seconds and say so.
Every claim is looked up
Each frame, stream, finding or file the answer cites is checked against the capture.
A citation that does not exist or says something false is marked, and the model is asked once to correct itself.
EnlargeSeven kinds of question
Each one has the same checks. Pick one to see what it gives you and where to find it.
Explain
A frame, a stream, an Expert finding, an Attack story item or an extracted file, in plain words.
The entries are in the menus and windows of frames, streams, expert findings, Attack story nodes and extracted files, and in the Detection window for a rule that fired.
Triage this capture
A first look at the whole capture, in four parts.
The hunting checks are run for it on your computer and put in the evidence, so the model starts from more than a summary.
- What happenedA few sentences on the capture as a whole.
- Look at firstA numbered list of at most seven things to open, each saying why and citing a frame, stream or finding.
- HostsWho took part, and what each one did.
- Not checkedThe model's own list of what it could not tell.
Chat
Any question about the capture, with the tools below.
Each question is sent with the tools it needs, not all of them. The answer says which tools it was based on. A Facts box takes notes the assistant should keep in mind, such as that an address is the file server or that a scanner is yours.
Describe the traffic
A display filter from a sentence.
The filter is parsed by Frame Bench and the frames are counted before anything is applied. If it does not parse, the model gets one correction with the parser's message. It is applied only when you press Apply.
EnlargeDraft a rule
A Suricata, Snort or YARA rule that was checked and tested.
The rule checker reads it, the engine installed on your computer runs it over the capture (after you say yes) and over a folder of ordinary traffic you choose, and the findings go back to the model for up to three revisions. Nothing is saved or loaded into a live engine by itself.
EnlargeReport text
A clearly labelled AI-written section for the Attack story report.
A switch leaves it out. The report names the model and the date.
Decode and explain
A script or an encoded file, peeled by Frame Bench and described by the model.
The layers are decoded with a fixed menu of steps (base64, hex, URL, inflate, character codes, XOR, reverse, UTF-16, white space). No code from the sample is ever run. The indicators the model names are searched for in the layers and marked Found or Not found.
EnlargeEvery answer is checked
A model can be wrong. Frame Bench looks up what it cites in the capture and says what it could not find.
- Found. The citation is a link to that frame, stream, finding, node or file, and a quoted value is compared with the real one.
- Not found. The line is folded under Not verified, and the model is asked once to correct itself.
- A number nobody gave. A note lists it. The note is advice and never changes the answer.
The download of /payload.exe was served by landing.exploit-kit.example F:26 and answered as an octet-stream F:27.
Not verified: A second upload to drop.exfil.example followed at F:412 and carried more…
It then contacted ads.redirector.example about every 30 seconds.
Numbers nobody gave. These numbers are not in the question, the evidence or any tool result: 30
3 claims cited, 2 verified, 1 not found. A made-up example, not a real answer.
Citations[F:123] for a frame, and streams (S:), findings (E:), Attack story nodes (N:), files (X:), hosts (H:) and alerts (A:).
A value can be quoted, as in [F:123 tcp.dstport=80], and is compared with the real one. A citation that checks out is a link to the frame, stream, finding, node or file.
What does not check out is markedA citation that does not exist, or says something false, is marked and the model is asked once to correct itself.
The first draft stays under Show original. When it cites a field the frame does not have, the model is told which fields that frame does have.
Lines that fail are foldedA line whose citation is not found is not shown as a finding.
It is folded under a Not verified line that begins with its first words. A note says how many were folded, and a heading is never folded.
Numbers nobody gaveNumbers in an answer that are in neither your question, the evidence nor any tool result are listed in a note.
Counts and totals should come from a tool; a number that is in none of them was probably worked out by the model. The note is advice and never changes the answer.
A first look before "not there"An answer that says something is not in the evidence, given without using a tool, is sent back once to check with the tools.
The evidence is a short summary and the tools read the whole capture.
Tools that keep failingCalls that fail twice in the same way, or three failed rounds, end the tool use and the model answers from what it has read.
A placeholder such as [result] or {id} is refused with a message that says to copy a real value.
A second opinionWhen the citations did not check out, a long answer cites nothing, or the model ran out of rounds, the panel offers the same question to another model you have set up.
You see what will be sent first, as always, and the earlier answer stays among the saved results.
What the check cannot doIt proves that what is cited is in the capture and says what the answer says it does.
It does not prove the conclusion is right. A fluent answer can still be wrong, so the answers are written as hypotheses: confirm them in the frames.
Nothing is sent before you have read it
What leaves the computer, and what is done to it first. Each level includes the one before.
Off
Nothing changes.
Secrets
Passwords, keys and tokens are replaced.
Default on this computerNames
Also addresses, host names and accounts, with stand-ins such as 198.18.0.7 or host-3.example.
Default for a cloud serviceStrict
Also leaves out the bytes of payloads.
Redaction lowers the risk and does not remove it. Read the preview. Lowering the level asks for a confirmation.
Off
10.0.2.15 > 203.0.113.30 ftp USER alice PASS hunter2
Secrets
10.0.2.15 > 203.0.113.30 ftp USER alice PASS [secret:password]
Names
198.18.0.7 > 198.18.0.8 ftp USER user-1 PASS [secret:password]
The same FTP login as the model would see it at each level. The stand-in addresses are examples. Your own screen shows the real values again, and Show the stand-ins lists what each one stands for.
What else keeps it in check
The exact text, firstThe preview shows the destination, the size, an estimated cost and what was replaced.
A model on this computer can be told not to show it again for the rest of that chat. A cloud service always shows it for the first question.
Stand-ins, and the real values put backThe answer is shown with your real values, and Show the stand-ins lists what each one stands for.
The table is kept in memory only. Accounts are read from the whole capture first, so a user name reaches the model as user-1 in every tool result and in your question.
Passwords in the clearRead from the whole capture first and told to the redactor.
FTP, POP3, IMAP, SMTP AUTH, HTTP Basic, Telnet, LDAP simple binds and form passwords are masked wherever they appear from the Secrets level up: in a frame, a stream, a hex dump or a list of strings, as text, as the base64 that carries them and as bytes.
Words in a packet cannot give ordersText from the capture reaches the model only inside blocks marked as untrusted data.
The list of tools does not change during a conversation, and answers are drawn as plain text, never as markup.
A record of every requestWhen, which model, where it went, how large and at which redaction level, with a CSV export.
The text of a request is kept only if you switch that on, and a key is never written to the log.
Your model, or your own cloud account
It is off until you set a model up in Preferences. Frame Bench talks to the model directly, and nothing goes through a service of ours.
On this computer
Ollama, LM Studio, llama.cpp or any server that speaks the OpenAI chat format.
A model counts as on this computer only if its address leads to this computer. There is no setting to type, so a cloud address cannot be passed off as local. A model on this computer is never sent through a proxy.
A cloud service, with your own key
OpenAI-compatible services and Anthropic.
Off until you turn it on in a window that says what leaves the computer. The address must be https. Offline mode stops all cloud contact at once, and a model on this computer keeps working. Frame Bench talks to the service directly: nothing goes through a service of ours.
Keys and headersAn API key is kept in the operating system's protected storage and is never shown again.
If the system has no protected storage, keys are held for this run only and Preferences says so. A gateway that wants a header of its own gets it from the model form, kept with the key and never shown again.
A model for each kind of questionExplain, triage, chat, the filter assistant, rule drafting, report text and decoding can each have their own.
A small model on this computer is enough for a filter or an explanation; a stronger one may be worth it for a triage or a report. If the model chosen for a kind of question is busy or missing, the saved model of the same service answers, once, and the answer says so.
Self-testThree questions about a small built-in capture, and a verdict: Good, Weak or Not usable for tools.
A model that cannot call the read-only tools is switched off for tools on its card and answers from the evidence pack alone. Test tries them again.
Context windowA model on this computer is asked how large its window really is.
When the size is left empty, the size the server reports is used instead of a careful 8192 tokens, and the preview says when the text does not fit.
ProxyCloud requests follow the computer's proxy settings, use none, or use one you enter.
HTTP, HTTPS, SOCKS4 or SOCKS5, with a list of addresses that skip it and a user name and password for a proxy that asks.
The tools it can use
A question is sent with the tools it needs, and the preview lists exactly which.
Tools in all
Most only read the capture, on your computer. A question is sent with the ones it needs.
Change what you see
Show a frame, add a note, add a star.
Ask you first
Nothing in this group runs without your yes. Adding frames it built can be allowed for a chat, or always.
Looking at the captureThe summary, frames, one frame decoded, bytes, searching, streams, hosts, protocol statistics and the filter parser.
Each result carries the ids it can be cited with, so a claim can be checked. A list of frames ends each line with what the application layer says, such as the Kerberos message and client, the DNS question or the HTTP request.
Hunting checksPeriodic contact, scans, DNS oddities, data leaving, logins in the clear, indicators, TLS, newcomers and capture quality.
quick_hunt runs eight of them in one call and host_profile stands in for a dozen. Results name what was found and what found nothing.
The frame toolboxEntropy, values, strings, checksums, byte operations, XOR, Base64, ciphers, decompression and templates.
Offered when the question points at bytes, or by name with /.
What it can changeThree tools change what you see or leave a mark: show a frame, a note, a star. Eight ask you first.
It cannot change the capture file on disk or a setting. With your yes it can change the bytes of one frame in the open capture, which you can undo, and save a template. It can send nothing onto a network unless you switch that on yourself, and every send is still confirmed. Adding frames the assistant built to the open capture asks once, and you can allow it for the rest of that chat, whether it is one packet or a whole stream; a new chat asks again. A switch in Preferences allows it always.
Offered by needBy default a question is sent with the core tools plus the groups its words point at.
Fewer still, for small models or a window of 16,384 tokens or less, and All of them, for a large model that should never have to ask. The preview lists exactly which tools go with the question, and the model can ask for more by name.
Costs, and keeping them small
Every call to a model costs tokens, and the same text goes again with each round. These are the ways Frame Bench trims it.
A ceiling and an estimateThe preview shows the estimated cost, the most the question could use if every allowed tool call is made, and what today has used.
Each model can have daily limits in tokens and in cost. A model with no price set shows no cost.
An allowance for each questionA question with tools makes several calls, and each sends the earlier text again. The most tokens one question may use is capped.
The default is 500,000 and the choices run from 100,000 to 2,000,000, or no limit. When it is reached the model is told to stop calling tools and answer from what it has read, and a note says the answer may be incomplete.
Only the tools a question needsThe description of every tool travels with every round, so a question carries the tools it needs.
The core tools are about 1,500 tokens. Small models get fewer.
Lists as tablesA list of records is sent as the column names once and then one row per record, about a third shorter.
A tool result is cut at 24 KB (less for a small window) and says how to narrow the request. Once an answer is finished, what the tools returned for it is not sent again.
The service's cacheWith Anthropic models the part of the request that does not change is marked for the service's cache.
The number marked cached under an answer shows how many tokens the service read back from it. The cache time is 5 minutes or 1 hour.
Short answers by defaultBrief is about 100 words. Normal and Detailed (up to about 500 words) are one choice away.
In Frame Bench's own evaluation (19 questions, several runs, two models) Brief kept the same recall of facts and the same valid citations as Normal with about a third fewer answer tokens.
A cheaper model for the reading stepsYour question goes to the main model first; a cheaper one makes the following tool calls.
The main model reads the same results and writes the answer. It is used only on the same side of the line as the main model, so a model on this computer never hands its data to a cloud service this way.
Saved with the case
Answers are kept, searchable, and tied to the evidence they cite.
Kept with the caseEvery answer is saved with the case and written into the session file, with a fingerprint of the evidence it cites.
When the session is opened again each result says Current, Evidence changed or Evidence missing. Always, Ask each time or Never is your choice.
The AI results windowSearch, filter, pin, annotate, copy, export, re-ask and delete.
Results can go into a new chat as earlier statements marked as not verified. Frames with results show an AI chip. A case with thousands of results opens at once.
Notes where you workThe first sentences of a kept result appear under the Expert finding or Attack story item it is about.
Citations stay clickable, and the note says so when the evidence it cites has changed.
Report a problemUnder every answer: what was wrong, and what it should have said.
The report keeps the question, the answer, the tools it used and the model, in the case file only. Nothing is sent anywhere. Reports can be exported as evaluation questions.
A chat belongs to its captureWhen you open another capture, the chat you had moves to History under the name of its capture.
Open the same file again, even under another name, and its History and its saved results come back by themselves, with no session file. Preferences can turn this off, or forget one capture or all of them.
The session file is as sensitive as the captureIt holds real addresses and names.
Save session without AI results leaves them out. Keys and the stand-in table are never saved.
EnlargeLimits to know about
What it needs, and what it cannot promise.
Small models are weakerSmall models on this computer write weaker explanations and sometimes ignore the citation rules.
The check marks what they get wrong, but a good answer needs a model of a few billion parameters or more and a context window big enough for the preview.
It reads what is decodedA capture with frames that are not loaded or not decoded gives it less to read.
Run Decode all frames first.
What is not visibleA login inside encrypted traffic is not visible to it, and it says how many frames it looked at.
The Not checked part of a triage is the model's own statement. Frame Bench does not check it, and a short list does not mean everything was checked.
Eight short videos
About 24 minutes in all, from the first question to the records. Each one is on its own subject.
Start here, 2:43From question to checked answer
The panel, the preview of what would be sent, redaction, citations looked up in the capture, the Facts box and the saved Results.
The videos use the real Frame Bench on a synthetic capture. The model on screen is a small stand-in on the recording computer with prepared answers, so that every take is the same, and each video says so. The preview, the redaction, the tools, the citation check and the saved results are the real ones. Nothing in them was sent to a cloud service. The video library asks you to sign in.
See the windows themselves.
The screenshots page has the panel, the preview, the results and the drafting windows at full size. The pricing table says which bundle holds the assistant.
All Rights Reserved © Netomize