AI Research Exposes Novel HTTP Desync Techniques and Apache Zero-Day

30,000 candidate desync vectors. That is how many ideas an AI-assisted system generated and tested while discovering novel HTTP desynchronization techniques and aiding a human researcher in finding an Apache Traffic Server zero-day, PortSwigger reported.

HTTP Terminator: how the system worked and what it scanned

PortSwigger said HTTP Terminator, an artificial intelligence–assisted research system built by James Kettle, ingested protocol text and generated large numbers of test cases. Kettle fed the system 138 HTTP and SMTP RFCs, split into roughly 15,000 small fragments that served as inspiration for 30,000 unique candidate vectors. PortSwigger reported the autonomous stage ran against 30,000 websites where scanning was authorized through bug bounty or vulnerability disclosure programs and initially flagged roughly 700 vulnerable targets before deeper validation and response-queue-poisoning (RQP) research.

New desynchronization techniques: dangling-byte and dual-matching Content-Length

The research produced several concrete technical outcomes. PortSwigger said the system generated new desync triggers and a dual-matching Content-Length pattern. An autonomous cascade tested 16 ideas aimed at improving the reliability of response queue poisoning; only a single method, dubbed the "dangling-byte" technique, survived evaluation. The dangling-byte approach deliberately leaves a smuggled request one byte short so that a second back-end response is not produced until a victim request supplies the missing byte, removing a race condition that previously made RQP unreliable on many sites.

Content-Type: multipart/byteranges finding and affected targets

One specific technique—using Content-Type: multipart/byteranges—worked across multiple server implementations and exposed more than 200 websites in the PortSwigger test set, the company said. Those findings included an unnamed U.S. bank, and Kettle said the broader discovery set involved banks, government infrastructure, security products, and an airport. PortSwigger emphasized that its scans were confined to targets where testing had been authorized.

Apache Traffic Server zero-day, disclosure status, and verification gap

PortSwigger also described a human-guided discovery cascade that led to a malformed request exposing a desynchronization zero-day in Apache Traffic Server. The researchers said that the issue has been patched and is tracked as CVE-2026-63078. However, an August 7 check by The Hacker News did not find a public record for CVE-2026-63078 on CVE.org or in the NVD, and Apache's July advisory covering 34 flaws did not list it. PortSwigger's report therefore leaves a verification gap: public records available at the time of reporting did not let defenders map CVE-2026-63078 to a specific fixed Traffic Server release.

Shared-Parser Confusion and the autonomy boundary

HTTP Terminator also proposed a broader concept PortSwigger calls Shared-Parser Confusion. Kettle said the system noticed that response-processing rules could be misapplied to requests when servers reuse parsing logic; the system proposed the idea, and Kettle validated and generalized it. As Kettle put it: "Neither of us would have discovered it alone." PortSwigger framed that distinction as the project’s autonomy boundary: the system autonomously generated and proved several techniques, while some findings, including the Apache zero-day and the Shared-Parser Confusion concept, required human intervention to validate and generalize.

Open-source tooling, models used, and related releases

PortSwigger has open-sourced HTTP Terminator. The paper does not map every autonomous discovery to a specific model or version: the released implementation uses Claude for document extraction and test-case generation, while its investigator stage requires Claude Code. Separately, researchers working on CRLF-powered desync attacks released public tools for studying that attack class—crlf-desyncs and crlf-powered-desync-scanner. Kettle also tested newer models on a rediscovery benchmark and reported a 30% success rate for GPT-5.6 Sol when given an inspiration technique.

What this means for technologists, affected enterprises, and policymakers

• Technologists and security teams: PortSwigger reiterated an operational defense—avoid HTTP/1.1 upstream where possible. Where HTTP/1.1 cannot be removed, it recommends allow-listing methods at both layers and restricting which methods may carry request bodies. Teams should also consider testing with the newly released tools to reproduce and validate desync behaviors in controlled environments.
• Affected enterprises and procurement leaders: the set of exposed targets included banks, government infrastructure, security products, and an airport, and the multipart/byteranges technique affected more than 200 sites in the test set. Enterprises will need to confirm whether their Traffic Server deployments were patched and to watch for a clear mapping of CVE-2026-63078 to a fixed release—something ported public records did not show at the time of reporting.
• Policymakers and regulators: the reporting highlights a practical coordination challenge—remediations tracked by vendors may not immediately appear in public CVE or NVD records, complicating external verification. That gap affects defenders who rely on public CVE mapping to prioritize and audit fixes.

The PortSwigger work illustrates two concurrent realities: AI-assisted systems can autonomously propose and validate protocol-level attacks at scale, and human researchers remain essential to validate, generalize, and shepherd fixes into vendors' supply chains. For defenders, the immediate tasks are familiar—apply patches where available, follow PortSwigger's HTTP/1.1 mitigation guidance, and seek clear CVE-to-release mappings for patched components—while keeping an eye on the newly public tools and techniques that will accelerate both defensive testing and adversary experimentation.

Original story — The Hacker News