| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| NetSentinel.exe | 2026-10-30 | 81.6 MB | |
| NetSentinel.exe.sha256 | 2026-10-30 | 83 Bytes | |
| NetSentinel v1.9.0 source code.tar.gz | 2026-10-30 | 988.7 kB | |
| NetSentinel v1.9.0 source code.zip | 2026-10-30 | 1.1 MB | |
| README.md | 2026-10-30 | 12.9 kB | |
| Totals: 5 Items | 83.6 MB | 0 | |
2100 unit tests (up from 1456) plus 14 integration checks. The suite no longer downloads threat feeds, so it runs in about a minute.
This PC's own traffic is inspected
The gateway, the DNS servers (8.8.8.8 and 1.1.1.1 included) and this computer's own addresses used to skip almost every rule. On a PC that mostly sees its own traffic, that was nearly all of it: port scans, DNS tunnels, uploads and gateway ARP spoofing on the machine NetSentinel protects were never looked at. Now only multicast, broadcast, link-local and discovery chatter, and devices you list as known, are skipped; the few rules that misfire on infrastructure exempt it themselves. Expect more alerts. Review mode is the way to sort them.
Also in live detection:
- Browsers have SNI and JA3/JA4 again. Post-quantum ClientHellos span two TCP segments and are now joined; QUIC Initial packets are decrypted for their server name; ports 8443 and 5061 are parsed.
- Capture files go through the live parser, on packet time: files get SNI, DNS answers and IPv6 ports, and replaying a capture no longer looks like a flood.
- Malware URLs are matched exactly on plain HTTP, live and in files.
- New detections: ARP spoofing with the gateway's MAC pinned across restarts; rogue DHCP, IPv6 router advertisements and DHCPv6; LLMNR, NBT-NS and mDNS poisoning (the Responder pattern); lateral movement over SMB, RDP, WinRM and RPC; uploads far above a host's own baseline; new devices, with their vendor; DNS tunnelling by length, volume and TXT rate; cryptocurrency mining pools; self-signed and expired server certificates; newly registered domains (WhoisDS's free daily list, checked locally).
- GeoIP, off until you turn it on: country and network operator of destinations from DB-IP Lite, downloaded monthly and looked up on this machine. IP Geolocation by DB-IP, CC BY 4.0.
- Fewer feed false alarms, the github.com pattern again: one alert per connection,
not one per direction; lists of attacking hosts no longer condemn cloud and CDN
addresses you connect to (shipped, dated ranges); delisted entries expire and a
week-old cache is not trusted; DShield's /24 blocks match; file hosts and URL
shorteners are never malicious domains; dynamic DNS matches on label boundaries
(
myfooddns.netis not a ddns.net host); fast-flux leaves CDNs alone. - Quieter: curl and okhttp are no longer CRITICAL; IOC-DGA, IOC-DOH-BYPASS and IOC-TOR stop repeating, and port 9001 alone is not Tor; IOC and ML alerts get the same cooldown as the IDS's.
- Bounded: forensics findings, the device table (now keyed by MAC) and capture buffers. A stop on a quiet interface really stops, and a restart no longer leaves the last run's capture threads going.
- URLhaus's keyless export is gone (HTTP 410). It now needs a free abuse.ch
Auth-Key (
threat_intel.abuse_ch_auth_key); abuse.ch's terms allow free not-for-profit use. A feed that answers 410 or 404 is not retried for a week.
Review mode: give Claude your alerts and the traffic behind them
--review DIR, or Settings → Review mode, keeps what it takes to judge alerts:
a capture of the traffic around every alert whatever its severity (cut before
verification can lower it), a rolling full capture within a disk budget, and
alerts.jsonl, written once each alert's verdict is final.
Build Review Bundle (desktop Review menu, the dashboard's Alerts page, or
python -m src.review build) writes a folder and a zip under 20 MB and 30 files:
a README telling the reviewer how to answer, a summary that points at rules firing
over and over on one host, the redacted alerts, and a text digest per group, one
line per packet with DNS names, TLS server names and HTTP host and path. Query
strings, cookies, credentials, tokens and user names in paths are never written;
LAN addresses can be replaced by labels; every file is audited for secret shapes
before it is written.
Verdicts come back the other way: Real or False positive on any alert in the dashboard, or import the JSON the README asks for (Review → Import verdicts). A false positive is never raised again for that rule and that host or domain.
An audit of your own network, when you ask
Settings → Network audit, then the Audit page: finds the devices on networks this
computer is on (ARP, or TCP without Npcap), scans 83 common ports at a limited
rate, and flags risky services (Telnet, FTP, SMB on workstations, RDP, VNC, UPnP,
WinRM, exposed databases, controller protocols, admin pages without TLS), each
with why it matters and how to fix it. It reads every TLS certificate, including
the TLS 1.3 ones the passive monitor cannot see, and compares each audit with the
last so new devices and newly opened ports stand out. Gentle mode (16 ports, 5 a
second) is for building and industrial controllers. It never tries a password,
never runs on its own, and only scans private networks of a /16 or smaller that
sit on one of this computer's interfaces. CLI: python -m src.audit --network
192.168.1.0/24. Its own probes are labelled as such in the alerts they cause.
File scans
- Secrets no longer leak through URLs in the report: query values, tokens before the @, token-shaped path segments, webhook paths and Azure SAS signatures are masked, and URLs end at a comma, so a browser password export cannot spill a password into the URL list. Browser and password-manager CSV exports are recognised and every saved password is reported masked.
- Weak feed guesses (dynamic DNS, a high-abuse TLD) are "suspicious" at LOW, not
known-bad at HIGH:
window.topno longer raises a HIGH alert. Internal names are never sent to threat feeds; UNC paths and internal host names are listed. - New secrets: Azure storage keys and SAS tokens, GCP service-account keys, Twilio,
SendGrid, Mailgun, npm and .npmrc, PyPI, GitLab, Discord and Slack webhooks,
Telegram bot tokens, Docker registry auths, bearer tokens, and
DB_PASSWORD=style settings and .env files. Windows leaks: unattend.xml passwords, Group Policy cpassword (decrypted and masked at once), Wi-Fi keys, KeePass databases, PuTTY keys. High-entropy values in settings. - Scripts: encoded PowerShell is decoded and scanned; download cradles and certutil, mshta, regsvr32, rundll32 and bitsadmin abuse are flagged, and Defender tampering. Office macros, .evtx event logs, and your own YARA rules if yara-python is installed.
- Archives by content: tar, gzip, bzip2, xz, 7z and zip-shaped files (.jar, .apk),
nested three deep, with a decompression budget and a zip-bomb check. RAR is
reported as not opened. Every file gets a SHA-256, checked against known-bad hash
lists you put in
~/.netsentinel/data/hash_lists(MalwareBazaar now requires a key even for its exports). - An ignore list for false positives, per value (by keyed hash) or, for cards and SSNs, per file and line. One alert per kind of secret per file. Folder scans run at background priority and give way between files.
Runs in the tray, and restarts itself
Closing the window used to stop monitoring and exit. Now it hides NetSentinel
to the system tray and monitoring carries on; the tray icon opens the window or
the dashboard, starts or stops monitoring, restarts, and quits (Settings →
Performance turns this off). Restart (in the window, the tray menu, and
offered after saving settings) starts a fresh copy with the same arguments and
resumes monitoring if it was running, so settings that apply at start take
effect without hunting for the process. --monitor starts monitoring as soon as
the window opens. A dashboard tab left open across a restart picks up the new
copy's token by itself, rather than answering every button with "missing or
wrong token" until it is reloaded. The tray uses pystray (LGPL-3.0) and Pillow.
The mouse wheel scrolls whatever is under the pointer. Three pages bound it for the whole window, so one turn scrolled all three, on whatever tab each was on, while the Incidents, ML, Settings and Forensics lists did not scroll at all.
What the first review of real alerts found
The first exe with review mode raised eight alerts in ten minutes, every one a false positive. Fixed:
- This computer was taken for a gateway. The default-route parser read every address on the line, the interface's included.
- A web server's version banner was "Sensitive Data in Transit". Plain-HTTP fetches from ESET, Windows Update, Let's Encrypt and Steam each raised a "[LIKELY THREAT]". Banners are now reported only for servers on your network.
- Signed software updates were insecure services. Windows Update, Edge's updater and others download over plain HTTP by design. A server stops counting as an updater the moment anything else is fetched from it.
- A WebRTC call was name poisoning. Browsers answer mDNS for random UUID names standing for their own addresses; those no longer count.
- The verifier called every sensitive-data finding "LIKELY THREAT 70%". It now goes by the kind of finding: card data, keys and control writes are confirmed; cookies and attack patterns are likely threats; banners, devices and uncommon HTTP clients (a headless browser run by a test, here) are left for you to judge.
- Review bundles built from the command line could not read the traffic. Every packet was "Raw". And when an alert involved this computer, the saved traffic was all of its traffic, and the alert's own connection fell past the digest's 400-line cut. Each digest now opens with the alert's own connection.
The dashboard can do things, on this computer only
Start and stop monitoring, acknowledge alerts, record verdicts, build review
bundles, run audits and manage the scan ignore list from the browser. These need
the token the dashboard page carries and are refused unless the server is bound
to localhost; configuration still cannot be changed over HTTP. The dashboard also
shows what NetSentinel costs (CPU, memory, threads, dropped and sampled-out
packets), and performance.low_priority runs it below normal priority.
netmon
- The web UI refuses foreign Host headers while bound to localhost, closing a DNS
rebinding path to its page token. Acknowledge, snooze and maintenance buttons
(
serve --allow-edit,web --state ... --allow-edit). - A model-triage button behind
--allow-summarise: it shows the exact redacted payload and an estimate first, and only a second click sends it. The daily budget is enforced and kept under--state, with an hourly cap and a timeout. Canary and configuration findings from metadata-only segments never reach the model. - The daily digest opens with an open-items checklist and marks acknowledged and maintenance findings; findings are kept across restarts and pushed once, not on every capture rotation; optional email digest over STARTTLS or SMTPS.
- GeoIP from DB-IP Lite (
netmon geoip --update); roleallowed_asnsenforced. - Inbound internet traffic is no longer dropped by router deduplication; new
inbound_https_bypassrule for origins published through Cloudflare. - Packet-side VLAN leak check: restricted frames on disk and restricted traffic
outside its VLAN are reported; the capture filter also excludes restricted
subnets and QinQ;
netmon bpf --verifyexits 3 on a leak. A capture filter with both host and port exclusions let untagged frames through; fixed.
Fixes
- Starting twice ran everything twice; a quick stop and start left the last run's loops going; a failed capture left the window saying MONITORING; the feature store stopped for good after a stop; alerts reached disk only on a clean stop.
- Alert captures matched only untagged Ethernet, so on a network that delivers VLAN-tagged frames an alert got a capture of everyone's traffic.
- The baseline whitelist learned by the last two labels, so one github.io or .co.za site whitelisted all of them; it learns registrable domains now.
- config.json saved by Notepad (UTF-8 with a BOM, or a non-ASCII name) failed to load and every setting fell back to defaults.
- The Settings threshold row edited a key the engine ignores; reloaded incidents said "0 alerts"; one incident kept every alert it absorbed.
- The elevated relaunch on Windows started in System32, so relative paths in arguments broke.
- CI's GitHub actions run on Node 24.
Installing. Install Npcap first (tick "WinPcap API-compatible
mode"), then run NetSentinel.exe. It asks for administrator rights because packet
capture needs them. NetSentinel.exe.sha256 holds the checksum to verify the download.