Analyzing Network Traffic

Analyzing Network Traffic

Packet capture is ground truth the endpoint can lie about but the wire cannot: every connection, DNS lookup, and byte transferred is recorded, whether or not the host's logs survived. The analysis is turning a flat capture into a story — who talked to whom, over what protocol, whether the pattern was human or automated, and what left the network. You are reconstructing intent from frames, not reading a verdict off a tool.

When to Use

• A .pcap / .pcapng capture needs forensic review for intrusion evidence
• A suspected C2 beacon must be confirmed and its interval, jitter, and channel characterized
• Data exfiltration is suspected and you need to size it, time it, and name the destination
• Malware network behaviour must be documented from the traffic it actually emitted
• A Zeek notice.log or Suricata eve.json alert needs to be run down to a verdict
• DNS tunneling, a DGA, or anomalous TLS (odd certs, rare JA3, SNI mismatch) is suspected

When NOT to Use

• Proactively sweeping endpoint, network, cloud, and identity telemetry for undetected compromise — use hunting-threats; this skill dissects one capture, a hunt spans data sources
• Running the wider incident — use responding-to-incidents; traffic analysis is one evidence stream feeding that process
• Detonating a sample to produce the traffic — use analyzing-malware; come here to analyze the pcap it emitted, not to run the binary
• The evidence is cloud control-plane activity, not packets (CloudTrail, VPC flow gaps, API calls) — use investigating-aws-incidents
• Actively testing a live application rather than analyzing a capture of it — use testing-web-applications

Capture and Handling

Get the capture right or every later step inherits the gap. A truncated snaplen or a dropped-packet capture cannot be fixed after the fact.

bashCopy code
# Full-frame capture, no name resolution, write to disk (never parse live)
tcpdump -i eth0 -nn -s 0 -w case.pcap
# -s 0 takes full frames; a default snaplen truncates payloads and breaks carving

# Ring buffer for long-running capture: 20 files of 200 MB, oldest recycled
tcpdump -i eth0 -nn -s 0 -w case-%Y%m%d-%H%M%S.pcap -G 3600 -C 200 -W 20

# Capture without dropping under load: raise the kernel buffer, filter tightly
tcpdump -i eth0 -nn -s 0 -B 4096 'not port 22' -w case.pcap
# Confirm drops after: the summary line reports "packets dropped by kernel"

Check whether the capture is intact and what you are holding:

bashCopy code
capinfos case.pcap        # packet count, duration, drop stats, snaplen, file type
tshark -r case.pcap -q -z io,phs   # protocol hierarchy — is the snaplen truncating?

• pcapng vs pcap: pcapng carries per-packet interface, comments, and name records; classic .pcap is a flat single-link format. editcap -F libpcap in.pcapng out.pcap downgrades for a tool that only reads pcap.
• Split a huge file so tools and memory cope: editcap -c 1000000 big.pcap chunk.pcap (per packet count) or editcap -i 600 big.pcap chunk.pcap (per 600 seconds).
• Merge related captures onto one timeline: mergecap -w all.pcap a.pcap b.pcap (sorts by timestamp).
• Trim to a window before shipping: editcap -A "2026-07-26 00:00:00" -B "2026-07-26 06:00:00" case.pcap window.pcap.
• Anonymize before sharing externally: tcprewrite/bittwiste to rewrite addresses, or editcap -s <snaplen> to truncate each packet to the first snaplen bytes (keeping headers, dropping trailing payload); TraceWrangler for deeper header/payload sanitization. Record what you changed so the recipient does not chase your rewrite as an artifact.

Hash the original and work on copies. sha256sum case.pcap goes in the case notes; the evidence file is read-only from here on.

The Triage Funnel

Start wide, narrow to the flows that matter. In Wireshark, work top-down:

• Statistics > Protocol Hierarchy — what protocols exist and in what proportion. Cleartext where you expected TLS, or a sliver of DNS carrying most of the bytes, is the first tell.
• Statistics > Conversations — sort by bytes and by duration. The longest and the largest flows are your first two leads (beacon vs bulk transfer).
• Statistics > Endpoints — which hosts talk to the most peers; map internal vs external at a glance.
• Follow > TCP/HTTP/TLS Stream — reassemble one conversation to read it as the application saw it.

Everything Wireshark does interactively, tshark does scriptably — which is how you extract fields across a whole capture instead of clicking:

bashCopy code
# Top talkers by bytes
tshark -r case.pcap -q -z conv,tcp

# Extract just the fields you want, tab-separated, for further processing
tshark -r case.pcap -T fields -E separator=/t \
  -e frame.time_epoch -e ip.src -e ip.dst -e tcp.dstport -e frame.len \
  -Y 'tcp.flags.syn==1 && tcp.flags.ack==0'   # every connection attempt

# Every HTTP request: host, method, URI, user-agent
tshark -r case.pcap -T fields -e http.host -e http.request.method \
  -e http.request.uri -e http.user_agent -Y http.request

Display-Filter Fluency

Display filters are the scalpel. The highest-value ones:

Chain them: ip.dst == 185.100.87.0/24 && dns scopes DNS to one suspect netblock. dns.qry.name matches "[a-f0-9]{20,}" flags long hex labels.

Zeek: the Workhorse

Zeek turns a pcap into structured, queryable logs — the single highest-leverage move on any capture bigger than a few thousand packets.

bashCopy code
zeek -r case.pcap
# Or add the community-id field for cross-tool pivoting:
zeek -r case.pcap policy/protocols/conn/community-id-logging
ls   # conn.log dns.log http.log ssl.log x509.log files.log notice.log weird.log ...

The log set and what each answers:

Mine them with zeek-cut (field extraction by name, order-independent):

bashCopy code
# Longest connections first — beacons and tunnels live at the top
cat conn.log | zeek-cut id.orig_h id.resp_h duration orig_bytes resp_bytes \
  | sort -t$'\t' -k3 -rn | head

# Every distinct destination one host reached, with hit counts
cat conn.log | zeek-cut id.orig_h id.resp_h | grep '^10\.0\.0\.5' \
  | sort | uniq -c | sort -rn

# DNS query names + types, to eyeball tunneling and DGA
cat dns.log | zeek-cut query qtype_name answers | sort | uniq -c | sort -rn | head -50

# Pivot a suspicious flow across all logs by its community-id
cat conn.log | zeek-cut community_id id.orig_h id.resp_h service

The community-id field is the same string across Zeek, Suricata, and many EDRs for the same flow — use it to line up an alert with the packets.

Suricata on a pcap

Run signatures offline against the capture to see what a rule set flags:

bashCopy code
# Run ET Open / community rules over the pcap, write structured events
suricata -r case.pcap -S /etc/suricata/rules/suricata.rules -l ./out/
# eve.json holds alerts, plus dns/http/tls/flow records if enabled

Extract and rank the alerts:

bashCopy code
jq -c 'select(.event_type=="alert") | {sig:.alert.signature, src:.src_ip, dst:.dest_ip}' \
  out/eve.json | sort | uniq -c | sort -rn

Read alerts as leads, not verdicts. A signature hit tells you where to look; it does not close the case, and its absence does not clear the capture.

Beacon Detection

Automated callbacks betray themselves in the timing and the shape, not the content. Look in conn.log for a source/destination pair that recurs at a regular interval, with small and roughly constant request sizes, over a long span. Human traffic is bursty and varied; a beacon is a metronome.

bashCopy code
# Inter-arrival deltas for one src/dst pair — near-constant gaps = beacon
tshark -r case.pcap -T fields -e frame.time_epoch \
  -Y 'ip.src==10.0.0.5 && ip.dst==185.100.87.202' \
  | awk 'NR>1{print $1-prev} {prev=$1}' | sort -n | uniq -c

• Regular intervals (e.g. every 60s) with jitter (a ±% spread) is the classic C2 signature — perfect regularity is rare malware, mild jitter is common.
• Small constant request sizes with occasional larger responses = check-in polling with tasking.
• Long connection duration or a flow that reconnects on a fixed cadence for hours points to a persistent channel.

Automate the scoring with RITA or a beacon-analysis tool over Zeek logs — they compute interval and size consistency across every pair so you are not eyeballing one at a time:

bashCopy code
rita import --database case --logs ./ && rita view case beacon:'>=90'

DNS Analysis

DNS is the covert channel of choice because it is rarely blocked and often unlogged. In dns.log, two distinct patterns:

Tunneling — DNS used as a data pipe:

• High query volume to a single parent domain (thousands of unique subdomains)
• Long labels: MFRGG43FMQ.tunnel.evil.com, base32/hex-looking subdomains
• TXT and NULL record types carrying encoded payload upstream/downstream
• High per-label entropy versus normal English-ish hostnames

bashCopy code
# Subdomain cardinality per parent domain — a tunnel spikes on one domain
cat dns.log | zeek-cut query | rev | cut -d. -f1-2 | rev \
  | sort | uniq -c | sort -rn | head

# TXT queries only — legitimate use is sparse; volume is a flag
cat dns.log | zeek-cut query qtype_name | awk -F'\t' '$2=="TXT"' | wc -l

DGA — algorithmically generated rendezvous domains:

• High NXDOMAIN rate (rcode 3) as the malware cycles through dead names
• Random-looking second-level domains with no linguistic structure

bashCopy code
cat dns.log | zeek-cut rcode_name query | awk -F'\t' '$1=="NXDOMAIN"' \
  | wc -l   # a burst of NXDOMAIN from one host is a DGA tell

Score label entropy or use a DGA classifier to separate cdn-3f2a.example (benign hash) from qwzjxkbvmn.info (generated). Volume plus entropy plus NXDOMAIN together make the case; any one alone has benign explanations.

TLS and Encrypted Traffic

You cannot read the plaintext without keys, but the handshake still fingerprints the client, the server, and the intent. Encryption hides content, not metadata.

• JA3 / JA3S (and JA4/JA4S) hash the ClientHello / ServerHello parameters into a client and server fingerprint. A rare JA3 shared across unrelated hosts, or a JA3 matching a known C2 framework, is a strong lead. Zeek's ssl.log carries ja3/ja3s; correlate against public and internal known-bad lists.
• In ssl.log / x509.log look for: self-signed certs, SNI that does not match the certificate CN/SAN, absurd validity windows (1000-year or same-day certs), empty/garbage subject fields, and SNI pointing at suspicious or newly-registered domains.

bashCopy code
# Self-signed or validation-failed TLS, with the SNI and JA3
cat ssl.log | zeek-cut server_name validation_status ja3 ja3s \
  | grep -iv '\bok$' | sort | uniq -c | sort -rn

# SNI vs certificate subject mismatch — join ssl.log and x509.log on cert id
cat x509.log | zeek-cut certificate.subject certificate.issuer \
  certificate.not_valid_before certificate.not_valid_after

When you legitimately hold the session keys (a lab detonation with SSLKEYLOGFILE set, or an exported master secret), decrypt in Wireshark: Preferences > Protocols > TLS > (Pre)-Master-Secret log filename, or tshark -r case.pcap -o tls.keylog_file:keys.log -Y http2. Never assume you can decrypt production TLS you have no keys for — you are fingerprinting, not reading.

HTTP and File Carving

Cleartext HTTP (and decrypted TLS) exposes the whole exchange:

• User-Agent anomalies — a hardcoded, malformed, or library-default UA (python-requests, an empty UA, a typo'd browser string) on outbound traffic is a common malware tell. Stack UAs and investigate the rare ones.
• Suspicious URIs in http.log — base64-looking paths, long random query strings, .php endpoints on a raw IP, or POSTs of opaque blobs.

Carve transferred files and hash them for pivoting:

bashCopy code
# Zeek extracts files automatically when configured; otherwise from files.log:
cat files.log | zeek-cut fuid mime_type filename md5 sha1 tx_hosts rx_hosts

# Carve without Zeek:
foremost -i case.pcap -o carved/          # signature-based file recovery
tcpflow -r case.pcap -o flows/            # reassemble every TCP stream to a file
# NetworkMiner (GUI/CLI) reassembles files, images, and credentials from a pcap

Hash every carved artifact and pivot suspicious ones to analyzing-malware: sha256sum carved/* — a file that appeared on the wire and matches nothing benign is the next sample to detonate.

Exfiltration Hunting

Data leaving is the outcome that matters most. In conn.log, sort by orig_bytes descending — large outbound flows are the headline, and the direction (orig = the internal host sending) is the whole point.

bashCopy code
# Biggest outbound transfers from internal hosts
cat conn.log | zeek-cut id.orig_h id.resp_h resp_p orig_bytes duration \
  | awk -F'\t' '$4>10000000' | sort -t$'\t' -k4 -rn

• Off-hours transfers — a bulk upload at 03:00 from a workstation that is idle by day. Cross the flow timestamps against business hours.
• Covert channels — steady ICMP with payload, or DNS carrying upstream data (see DNS analysis). Small packets, high count, one destination.
• Destination reputation — cloud storage (*.s3.amazonaws.com, *.blob.core.windows.net), paste sites (pastebin, ghostbin), and file-share domains as the destination of a large upload from a server that has no business reason to use them.

Protocol Anomalies

Attackers hide traffic on the wrong port and in the wrong protocol:

• Non-standard ports — TLS on 8443 or 4443, HTTP on 8080/8888, or C2 on a high random port. conn.log's service field is Zeek's detected protocol, independent of port number.
• Protocol-on-wrong-port — Zeek's DPI flags SSH on 443, HTTP on 53, or TLS on 22 in conn.log (service disagrees with id.resp_p) and in weird.log.

bashCopy code
# Detected service does not match the port — tunneling / evasion
cat conn.log | zeek-cut id.resp_p service | awk -F'\t' \
  '($1=="443" && $2!="ssl") || ($1=="53" && $2!="dns")' | sort | uniq -c

• Cleartext credentials — FTP, Telnet, HTTP Basic, SMTP AUTH, LDAP simple bind. frame contains "PASS " in Wireshark, or NetworkMiner's credentials tab, pulls them straight out.

Handing Off IOCs

The capture's value is what you extract for reuse. Tier indicators the way you would from any source — behaviour outlives infrastructure — and route each output:

• File signatures from carved samples — write .yar in writing-yara-rules against structure, not the URI it happened to use.
• Packaging IOCs, actor pivots, and infrastructure clustering into a finished product — producing-threat-intelligence.
• Turning a traffic pattern into a deployed detection — writing-sigma-rules for log-based rules and engineering-detections for the broader rule pipeline (a Suricata signature for the JA3, a Zeek notice for the beacon cadence, a Sigma rule for the DNS pattern).
• The finished write-up for stakeholders — reporting-security-findings.

Record for every indicator: the flow it came from, the timestamp, and your confidence. An IP with no context is noise to whoever receives it.

Rationalizations to Reject

• "It's all TLS, I can't see anything." JA3/JA3S, SNI, certificate fields, and connection timing fingerprint encrypted traffic without decrypting it. The metadata is the analysis.
• "No alerts fired, so it's clean." Beaconing and DNS tunneling routinely slip past default rule sets. Absence of a Suricata hit is not absence of C2.
• "The pcap is too big to analyze." Run Zeek, work the logs, and filter down. You triage structured logs, not five million raw frames.
• "There's no C2 traffic here." Check DNS, ICMP, and long-lived low-volume flows before concluding. C2 hides in the channels you did not sort by bytes.
• "The volume is normal, so no exfil." Sort by outbound bytes and by time. A slow drip over days, or one off-hours burst, does not move the average.
• "That destination is just a CDN / cloud provider." Cloud storage and CDN domains are exactly where modern exfil and C2 hide. Confirm the flow's shape, do not wave it through on the domain name.
• "I'll just look at the payloads." Truncated snaplen, encryption, and reassembly gaps mean payload-first misses the story. Start with the flow metadata, then read the streams that earned it.

ATT&CK Coverage

Generated from secskills-core/ttp-index.json — edit that file, then run python3 scripts/sync_attack.py --write. Re-verify IDs against the current ATT&CK release before citing them in a report.

Command and Control (TA0011)

• T1071 Application Layer Protocol — see also analyzing-malware, engineering-detections
• T1071.004 DNS — see also engineering-detections, hunting-threats
• T1132 Data Encoding — see also transferring-files, analyzing-malware
• T1568 Dynamic Resolution — see also analyzing-malware, hunting-threats
• T1573 Encrypted Channel — see also analyzing-malware, engineering-detections

Exfiltration (TA0010)

• T1048 Exfiltration Over Alternative Protocol — see also transferring-files, hunting-threats

Detection content for any of these: engineering-detections. Proactive search: hunting-threats. Post-compromise: responding-to-incidents.

References

• hunting-threats — proactive, hypothesis-driven search across many telemetry sources, of which network data is one
• responding-to-incidents — the incident process this evidence stream feeds
• analyzing-malware — detonate the sample to produce traffic, and receive carved files from this skill for analysis
• investigating-aws-incidents — when the evidence is cloud control-plane logs rather than packets
• writing-yara-rules — file signatures from carved artifacts
• producing-threat-intelligence — packaging extracted IOCs into a product
• reporting-security-findings — the stakeholder write-up
• Wireshark / tshark and capinfos — interactive and scripted dissection
• Zeek with zeek-cut and community-id — structured logs from any capture
• Suricata (ET Open / community rules) — offline signature runs over a pcap
• NetworkMiner — file, image, and credential reassembly from captures
• RITA / beacon-analysis — automated beacon scoring over Zeek logs
• tcpdump — capture; editcap / mergecap — split, merge, trim, and convert