Most conversations about WAFs treat them as a monolithic thing you bolt onto a network edge and forget. But the word “WAF” gets applied to tools operating at very different depths in the stack. A product that rate-limits TCP connections and one that parses every HTTP header are doing fundamentally different jobs, even if both call themselves application firewalls. The OSI layer at which inspection happens determines what an attacker can hide from you.
The OSI Stack and Where Inspection Happens
The OSI model gives each layer a specific job. A firewall operating at Layer 3 sees IP packets. One at Layer 4 sees TCP/UDP segments. At Layer 5, you see sessions: connection establishment, TLS handshakes, and transport-level state. At Layer 7, you see the actual HTTP request, including its method, path, headers, cookies, and body. That difference in vantage point is not cosmetic.
OSI Model: WAF Inspection Points
L7 Application │ HTTP/HTTPS, GraphQL, WebSockets: full request payload visible │ ← L7 WAF
L6 Presentation │ TLS encryption/decryption, encoding negotiation │ TLS offload
L5 Session │ Connection state, session tokens, TLS handshake, SYN patterns │ ← Session FW
L4 Transport │ TCP/UDP ports, segment reassembly, rate limiting │ NGFW / ACL
L3 Network │ IP routing, geo-blocking, IP reputation lists │ IP FW / CDN
What a Session-Layer (L5) Firewall Sees
Session-layer firewalls work before HTTP ever gets parsed. They track whether a TCP three-way handshake completed, how fast connections are arriving, how long TLS negotiation takes, and whether a session token is being reused across unexpected IP ranges. What they cannot see is the payload once the session is established.
This makes them very effective against volumetric attacks and connection-state exhaustion. A SYN flood, a TLS fingerprint anomaly, or a slow-open attack is detectable here without touching application logic. A SQL injection string inside a POST body is invisible.
# Session-layer firewall log: connection events
2026-10-26T04:31:12Z CONN 203.0.113.44:52111 → 10.0.0.1:443 SYN ALLOW
2026-10-26T04:31:12Z CONN 203.0.113.44:52112 → 10.0.0.1:443 SYN ALLOW
2026-10-26T04:31:12Z CONN 203.0.113.44:52113 → 10.0.0.1:443 SYN ALLOW
...
2026-10-26T04:31:12Z ALERT 203.0.113.44 → 10.0.0.1:443 SYN_FLOOD rate=4200/s > threshold=500/s
2026-10-26T04:31:12Z DROP 203.0.113.44/32 → any:443 BLOCK ttl=300s
# L5 knows: connection rate, TLS fingerprint, session state
# L5 blind to: HTTP method, URL path, POST body, cookies
What an Application-Layer (L7) WAF Sees
A Layer 7 WAF terminates TLS, reassembles TCP streams, and reads the full HTTP request before deciding what to do with it. It can inspect query string parameters, request bodies, cookie values, custom headers, and JSON payloads. OWASP Top 10 attacks, XXE, SSRF, and most injection vulnerabilities live entirely in this layer. No amount of connection-state analysis catches them.
The trade-off is computational cost and latency. Deep packet inspection at L7 is orders of magnitude more expensive per connection than a session-state table lookup. A volumetric DDoS carrying millions of connections per second will exhaust a cloud WAF’s inspection budget long before it reaches rules that check for SQLi patterns.
# L7 WAF log: HTTP inspection events
2026-10-26T04:31:58Z BLOCK 198.51.100.7
method POST
path /api/v2/users/search
body username=admin'--+OR+1%3D1--
rule SQLI-942100 SQLi via boolean-based blind
score 90/100 BLOCK
2026-10-26T04:32:11Z BLOCK 198.51.100.12
method GET
path /etc/passwd
header X-Forwarded-For: 127.0.0.1/../../../etc/passwd
rule LFI-930110 Path traversal in header
score 95/100 BLOCK
# L7 sees everything the L5 firewall passes through unchecked
A Direct Comparison
L3-L5 Session Firewall
- Handles millions of connections per second
- Stops SYN floods, UDP floods, slow-open attacks
- TLS fingerprint anomaly detection (JA3/JA4)
- IP reputation and geo-blocking at wire speed
- Cannot see HTTP method, path, or body
- Blind to injection, XSS, SSRF
- Cannot enforce rate limits per endpoint
L7 Application WAF
- Full HTTP inspection: method, path, headers, body
- OWASP Top 10, XXE, SSRF, IDOR detection
- Per-endpoint rate limiting and bot scoring
- Schema validation for API payloads
- Expensive per connection; overwhelmed by floods
- Requires TLS termination (introduces a trust boundary)
- High false-positive rate without tuning
Attack Coverage Matrix
| Attack Type | L3-L5 FW | L7 WAF | Notes |
|---|---|---|---|
| SYN / UDP volumetric flood | Blocks | Overwhelmed | L3-L5 only |
| SQL injection (POST body) | Invisible | Blocks | L7 only |
| Reflected XSS in query param | Invisible | Blocks | L7 only |
| Slow-read / slow-POST attack | Detects | Partial | L3-L5 primary |
| TLS downgrade / abnormal handshake | Detects | Partial | L3-L5 primary |
| API schema violation (JSON body) | Invisible | Blocks | L7 only |
| L7 HTTP flood (high RPS per IP) | Rate limits IP | Per-endpoint | Both complement |
| Bot traffic with valid TLS fingerprint | Passes | Scores and blocks | L7 only |
The Real-World Architecture
Production deployments do not choose between these layers. A CDN edge scrubs volumetric traffic at L3-L4 before it hits your origin. A session-aware layer (L5) applies connection throttling and TLS anomaly detection. A WAF at L7 inspects every request that survives those two filters. Each layer handles what the next one cannot afford to process.
The key point: an L7 WAF that receives a DDoS flood will spend its CPU budget parsing malformed SYN packets instead of inspecting requests. An L5 firewall blocking by IP will pass a slow SQLi campaign from a residential proxy pool. Stacking the layers is not redundancy. It is defence-in-depth with each layer doing the job it is computationally suited for.
Vendor Terminology is Inconsistent
One more thing worth knowing: vendors do not agree on layer labels. Some products marketed as “WAFs” operate primarily at L4-L5 with basic signature matching. Others marketed as “next-generation firewalls” do full L7 DPI. Reading vendor marketing through the OSI lens cuts through the noise. Ask the vendor what they terminate, what they parse, and what they cannot see, and you will get a clearer picture of what you are actually buying than any product name provides.
The OSI model is a simplification of reality, but it is a useful one here. The row where your firewall does its work determines its fundamental capabilities. Everything else is implementation detail.