AI-Assisted HTTP Terminator Finds Novel HTTP Desync Techniques and Apache Zero-Day

What it is

The page carries specifics the catalog summary omits: a concrete defense, a named request-parsing technique with a real-world hit count, the mechanism of the dangling-byte trick, the target sectors hit, and a model-benchmark figure. HTTP Terminator (PortSwigger, James Kettle) is already described in the summary — what follows is what the summary cannot.

How it works

The Content-Type: multipart/byteranges technique worked across multiple server implementations and exposed more than 200 websites in the test set, including an unnamed U.S. bank. This is the single highest-yield vector the page names.

The autonomous run tested 16 ideas for improving response queue poisoning (RQP); only the dangling-byte technique survived. It leaves a smuggled request one byte short, so the second back-end response is not produced until a victim request supplies the missing byte — eliminating the race condition that otherwise makes RQP unreliable across many sites. RQP itself can make a front end lose track of which back-end response belongs to which user, exposing another user's response including session cookies or API keys.

Shared-Parser Confusion arises when servers reuse parsing logic and response-processing rules get misapplied to requests.

Preconditions / Affected

Vulnerable hosts run HTTP/1.1 upstream (front end to back end). Kettle said confirmed findings involved banks, government infrastructure, security products, and an airport — i.e. this is live across high-value sectors, not a lab curiosity.

What to do

The defense the page states, in priority order:

• Avoid HTTP/1.1 upstream. Remove HTTP/1.1 on the front-end-to-back-end hop where possible.
• Where HTTP/1.1 cannot be removed, allow-list methods at both layers.
• Restrict which methods may carry request bodies.

Caveats

• CVE-2026-63078 (Apache Traffic Server): an August 7 check by The Hacker News found no public record in CVE.org or NVD, and Apache's July advisory covering 34 flaws did not list it. Defenders cannot yet map the CVE to a specific fixed Traffic Server release from public records.
• The Apache zero-day and Shared-Parser Confusion required Kettle's human intervention; only the desync triggers, dual-matching Content-Length pattern, and dangling-byte were proved autonomously. The paper does not identify which model or version produced each autonomous discovery.
• On a separate rediscovery benchmark, Kettle reported a 30% success rate for GPT-5.6 Sol when given an inspiration technique.