PortSwigger AI System Uncovers HTTP Desync Methods and Apache Flaw
PortSwigger says its AI-assisted HTTP Terminator research system generated and proved new HTTP desynchronization techniques after evaluating 30,000 candidate attack vectors. The project, built by PortSwigger research director James Kettle, also contributed through a separate human-guided research cascade to discovery of an Apache Traffic Server zero-day tracked as CVE-2026-63078.
Kettle said HTTP Terminator scanned 30,000 websites where testing was authorized through bug bounty or vulnerability disclosure programmes. It identified roughly 700 vulnerable targets before deeper validation and response queue poisoning research. Those targets included banks, government infrastructure, security products and an airport.
New ways to trigger desynchronization
HTTP request desynchronization arises when front-end and back-end systems interpret the boundaries of an HTTP request differently. The research produced new desync triggers, a dual-matching Content-Length pattern and a dangling-byte technique intended to improve the reliability of response queue poisoning, or RQP.
In an RQP scenario, a front end can lose track of which back-end response belongs to which user. That can potentially expose another user’s response, including session cookies or API keys. HTTP Terminator tested 16 ideas for improving RQP; the dangling-byte technique was the only one that survived evaluation.
The technique leaves a smuggled request one byte short. The second back-end response is then not produced until a victim request supplies the missing byte, removing a race condition that makes RQP unreliable on many sites. The broader desynchronization threat also features in AI-related attacks and data exposure risks, where AI-related attacks and data exposure were among the security issues examined.
How the system was built and where human validation remained necessary
Kettle fed the system 138 HTTP and SMTP RFCs, split into about 15,000 small fragments. These fragments served as inspiration for 30,000 unique candidate vectors. One technique using Content-Type: multipart/byteranges worked across multiple server implementations and exposed more than 200 websites in the test set, including an unnamed US bank.
The researchers also described Shared-Parser Confusion, an attack concept proposed by HTTP Terminator after it identified that response-processing rules could be applied to requests when servers reuse parsing logic. Kettle validated and generalized the idea. He said neither the system nor the researcher would have discovered it alone.
The Apache Traffic Server issue emerged from a human-guided cascade involving a malformed request. PortSwigger said the issue has been patched, but an August 7 check did not find a public CVE.org or NVD record for CVE-2026-63078, and Apache’s July advisory covering 34 flaws did not list it. Public records therefore did not identify the specific fixed Traffic Server release.
Defensive priorities
PortSwigger has open-sourced HTTP Terminator. Its released implementation uses Claude for document extraction and test-case generation, while the investigator stage requires Claude Code. The paper does not identify the exact model or version responsible for each autonomous discovery.
PortSwigger’s mitigation advice remains to avoid HTTP/1.1 upstream where possible. Where it cannot be removed, organisations should allow-list methods at both layers and limit request bodies to the methods that need them. For businesses operating layered web infrastructure, reviewing those controls is a practical response while release-specific remediation details for the Apache case remain unavailable.