Hunt Smuggling Agent Skill

Hunt: HTTP Request Smuggling

Assumes hunt-core for the scope gate, two-account rule, confirmation gate, enumeration limits, stop conditions, wiki protocol, FIND output, and Deadends. Do not re-derive any of that here.

Wiki

qmd_query "HTTP request smuggling desync CL.TE TE.CL TE.TE CL.0 HTTP/2 downgrade" via wiki-search MCP

Hub: [[web-moc]] (live web index). Primary page: [[http-request-smuggling]]. Payload arsenal: wiki/payloads/smuggling.md.

Anchors: [[reverse-proxy-attacks]].

Attack surface signals

Smuggling needs a front-end that parses headers differently from the back-end: a CDN/WAF/LB in

front of an origin, HTTP/1.1 keep-alive reused across users, or an HTTP/2 edge that downgrades to

HTTP/1.1. Odds rise wherever both Transfer-Encoding and Content-Length are honored somewhere

in the chain.

Rank before probing. Not all front-end/back-end pairs are equally likely:

• HTTP/2 edge downgrading to HTTP/1.1 origin - H2.CL / H2.TE re-open a surface a hardened H1

front-end closes; highest yield on modern stacks.

• Multiple hops in the chain - CDN -> WAF -> LB -> origin; each added parser is another chance

for two of them to disagree.

• Mismatched server software front vs back - e.g. a normalising proxy in front of a lenient

app server, or vice versa.

• Connection reuse to the back-end - a poisoned prefix only reaches a victim if the back-end

pools/keep-alives the connection across requests.

• Endpoints behind the edge auth/path filter - a desync that bypasses a front-end control is

worth more than one that does not reach anything privileged.

Methodology

• Detect (timing, safe): send a deliberately malformed TE/CL and watch for a back-end read timeout.

# CL.TE probe (front-end uses CL, back-end uses TE) - delays if vulnerable
POST / HTTP/1.1
Content-Length: 4
Transfer-Encoding: chunked

1
A
0

# TE.CL probe (front-end TE, back-end CL)
Transfer-Encoding: chunked  +  Content-Length: 6 ; body: "0\r\n\r\nX"

• Confirm (differential): smuggle a prefix that prepends to the victim's request, then issue a normal request and observe the poisoned response (e.g. your G prepended to their path -> 404 on GET ...). See the confirmation gate below - a timing delay alone is never the finding.
• TE.TE: obfuscate the header so one server ignores it (Transfer-Encoding: xchunked, Transfer-Encoding, Transfer-Encoding:\tchunked, double TE).
• HTTP/2: test H2.CL / H2.TE (smuggle via H2 that downgrades to H1), and H2 request splitting via CRLF in header values.
• CL.0 / H2.0: back-end ignores body -> smuggle a full second request.
• Exploit: capture other users' requests (steal session cookies/headers), bypass front-end auth/path controls, cache-poison via smuggled response, escalate a reflected issue to stored.
• Distill when confirmed - reusable obfuscation or H2-desync variant, GENERIC, no client host: python3 scripts/wiki-stage.py --kind technique --slug <slug> --target-page techniques/web/http-request-smuggling.md

Drive it through Burp for operator visibility. Use Burp Repeater "Send group in sequence

(single connection)" for the differential and the HTTP Request Smuggler extension for the sweep;

push each load-bearing smuggled request into Repeater rather than leaving it in a curl loop.

Confirmation gate

NOT confirmation: a single slow response or timing noise alone; a back-end read timeout by

itself; a malformed request the front-end simply rejects; a delay you have not reproduced. A

single timing observation is a signal to investigate, never proof of a desync.

IS confirmation: a reliable differential - a smuggled prefix that changes the *next* request's

response (your own follow-up, or a captured victim request), reproduced on a clean connection; or

a consistent timing delta reproduced across several runs and matched to a specific desync variant;

or a DNS/HTTP hit to your unique Burp Collaborator / interactsh subdomain fired by a

Collaborator-pointed smuggled request.

Blind / OOB desync. When you plant a Collaborator-pointed smuggled request, append a row to

targets/<eng>/oob.md: | <token> | <sink url+param> | smuggling | <date> | waiting | |

(columns: token | sink | class | planted | status | source; token = your unique Burp

Collaborator / interactsh label). The recon-capture hook auto-correlates the incoming callback to

flip the row to HIT and SessionStart surfaces HITs; a HIT row is the gate to scaffold the FIND.

Do NOT claim a blind smuggling desync without the captured differential or a HIT row.

Chaining

A confirmed desync is a delivery primitive, not the endpoint. Chain it:

• Smuggled response -> cache poisoning. Poison a shared cache with a smuggled response so it

serves to every user; hand off to hunt-cache.

• Request hijack. Capture the next user's full request (session cookie, auth header, CSRF

token) by smuggling a prefix that appends their request to a body you read back.

• Front-end control bypass -> stored. Smuggle past an edge auth/path filter to reach an

internal endpoint, or escalate a reflected issue to stored via a poisoned response.

Evasion

If the direct CL/TE probe is normalised clean, the desync is often still there behind

obfuscation - work the TE.TE variants (step 3) and the HTTP/2 downgrade (step 4), which re-opens

the surface a hardened HTTP/1.1 front-end closes. Header casing, line-folding, and bare-CR/LF

tricks beyond those live in wiki/payloads/smuggling.md.

Severity

| Impact | Typical |

|---|---|

| Captured victim requests (session cookie / auth header) at scale | critical |

| Admin session theft | critical |

| Front-end auth / path-control bypass | high |

| Cache poisoning via smuggled response | high |

| Reflected escalated to stored via poisoning | high |

Rated on demonstrated impact per hunt-core. A desync you can trigger but cannot yet show

capturing a victim request or poisoning a response is a mechanism, not the severity ceiling.

Deadends

Append: - [ ] smuggling on <host> -- CL.TE/TE.CL/TE.TE/CL.0/H2 all clean
              (single normalising front-end, no timing delta or differential after a full sweep)

Record which variants you swept, not just that it failed. The next pass needs to know the boundary.