SonicWall appliances guard thousands of professional-services offices, the exact environments where a dictation app with silent training terms is a genuine liability. The syslog those boxes already emit is the audit's raw material.
SonicWall shops usually mean a single admin carrying security alongside everything else. This loop is sized accordingly.
Pull a syslog slice from wherever it already lands, upload, read. No new infrastructure, no agents on partner laptops, no change-control meeting.
When a managing partner asks "are we exposed?", the answer becomes a dated PDF instead of a hedge. That difference is the whole product.
The syslog server you stood up for compliance years ago finally earns its keep: the audit reads raw key=value lines exactly as SonicWall writes them.
Web-filter and connection events carry the four values that matter. A realistic slice, noise dimmed:
A 45-lawyer firm, one SonicWall, one IT manager, SSO agent on the domain. The trigger was an e-discovery vendor asking about AI use. Sample data, brief format.
The IT manager exports 30 days of web-filter syslog: 150,000 lines. Upload, match, 16 AI tools found. Twelve are unremarkable. Four are dictation or summarization tools, which in a law firm means client-confidential material is the input.
The per-user table shows two dictation apps concentrated among four associates, plus a summarizer being fed briefs by a paralegal team. One tool trains unless opted out; two say nothing about training at all.
Our wider terms analysis says silence is the norm, 85.5% of AI tools state nothing about training. Law firms cannot file "the norm" as due diligence.
The firm sanctions one legal-grade dictation vendor after a terms review, blocks the consumer tools at the appliance, and adds an AI clause to the staff handbook.
The audit PDF goes into the e-discovery vendor response and the professional-indemnity renewal file, both of which had asked the same question.
One $99 report, roughly ninety minutes of the IT manager's time, and a partner meeting that ran short because the tiles page needed no explanation.
The usr= field depends on the SSO agent. Every state still yields a valid inventory.
Names on every line for domain-joined machines. The per-user table supports the exact conversation the law firm needed.
VPN users, terminal servers and non-domain devices may log empty usr=. Their src IPs still count as sources, so nothing disappears.
Whole export attributes by IP. Pair the report with your DHCP reservations and the reading is nearly as sharp.
| Report section | With SonicWall syslog |
|---|---|
| Summary tiles | Full totals: tools, high-risk, training exposure, abusive-purpose, unsanctioned. |
| Tool table | Every matched dstname with category, risk, sovereignty and dated training verdicts. |
| Per-user table | usr= identities plus IP-only sources, side by side. |
| Sanctioned split | The firm's approved list against observed use, the professional-duty number. |
| Blocked evidence | CFS-blocked events stay visible as enforcement proof if included in the export. |
| CSV + PDF | Full-report tier: the PDF for partners and insurers, the CSV for the risk register. |
Uploads are read once and discarded; arg= paths never reach the report. Reports live 90 days and can be deleted earlier, which belongs in the firm's own data-handling note.
The sample evidence pack shows the full document; the free preview shows the same with masked rows you can count. Both are one click from here.
The law firm's pattern generalizes to accountants, advisors and consultancies: confidentiality is the product, and AI tools are where it leaks quietly.
Dictation, summarization and drafting tools all ingest client material by design. The report's training verdicts, dated, are the due-diligence artifact for that exposure.
Professional-indemnity questionnaires increasingly include AI-usage questions. A log-based inventory answers in a sentence plus an attachment.
E-discovery vendors, corporate clients and outside counsel guidelines all probe AI use now. Firms with a quarterly audit answer from the file, not from memory. The evidence-chain angle continues on the compliance officer page.
Quarterly is the professional-services cadence: one report per quarter, filed with the matter-management or risk system. Plans and packs are on the pricing page.
Web-filter and connection events carrying dstname=, as raw key=value syslog. No reformatting; collector slices upload as-is.
Yes. Attribution falls back to source IPs, and the tool inventory is identical. DHCP reservations make the IP table readable.
No. Hostnames arrive via SNI in dstname= regardless, and the audit needs nothing deeper.
It should. Blocked AI events document enforcement, and attempt volume against blocks signals demand for a sanctioned alternative.
The audit reads hostnames, identities and timestamps; arg= paths are dropped at parse. The upload is discarded after the run and reports are deletable early.
The law firm took about ninety minutes including reading time. The export is the slow part; matching takes minutes.
From the law firm's report, sample data. This one table justified the audit to the partners in a glance.
| Tool found | CFS category at the time | Audit verdict |
|---|---|---|
| Dictation app A | Business | Block: trains unless opted out, audio input |
| Summarizer | Education | Control: terms silent on training |
| Chatbot | Information Technology | Control: enterprise tier acceptable |
| Legal-grade dictation | Business | Allow: contractual no-training terms |
Same CFS category, opposite verdicts, twice. Category filtering cannot make that distinction; per-tool verdicts with dated terms checks can.
The firm's handbook addition was short because the audit does the enforcement measuring. A pattern worth borrowing:
"Client material may only be entered into tools on the approved AI list." One sentence, no technology lecture, enforceable because the list exists.
"Requests for new tools go to IT and are answered within two business days." Fast sanctioning is what keeps associates from not asking.
"Usage is reviewed quarterly from network logs." The audit is the review; the sanctioned split is the metric; nobody is surprised later.
Building the approved list itself, tier by tier, is covered on sanctioned vs unsanctioned AI.
Thirty days of syslog, one upload, and "are we exposed?" gets a dated, specific, fileable answer.
Run the free audit