**PortSwigger**'s **HTTP Terminator**, an AI-assisted research system spearheaded by **James Kettle**, has made significant strides in uncovering novel HTTP desynchronization vulnerabilities. By exploring 30,000 candidate desync vectors, the system autonomously generated and validated new attack techniques. ### AI-Driven Vulnerability Discovery **Kettle** reported that **HTTP Terminator** tested 30,000 websites, authorized through bug bounty or vulnerability disclosure programs, identifying approximately 700 vulnerable targets. These findings spanned critical infrastructure, including banks, government systems, security products, and an airport. ### New Desync Triggers and RQP Enhancements The research yielded new desynchronization triggers, including a dual-matching Content-Length pattern and a novel "dangling-byte" technique. The dangling-byte method significantly enhances the reliability of Response Queue Poisoning (RQP), an attack that can cause a front-end server to misdirect back-end responses, potentially exposing sensitive data like session cookies or API keys to unintended users. ### Shared-Parser Confusion: A Broader Attack Concept **HTTP Terminator** also proposed the concept of Shared-Parser Confusion, a broader attack where servers reuse parsing logic, leading to the misapplication of response-processing rules to requests. While the system proposed this concept, **Kettle** validated and generalized it, highlighting the collaborative synergy between AI and human expertise. ### Addressing the Apache Zero-Day During a separate human-guided discovery cascade, a malformed request exposed a zero-day desynchronization vulnerability in **Apache Traffic Server**. This issue has reportedly been patched and assigned **CVE-2026-63078**. However, public records for this CVE were not immediately available on CVE.org or NVD, leaving a verification gap for defenders. ### Mitigation Strategies Remain Crucial Despite these new discoveries, the core defense recommendations against HTTP desynchronization remain consistent. **PortSwigger** advises avoiding **HTTP/1.1** upstream where possible. If **HTTP/1.1** cannot be removed, it recommends allow-listing methods at both layers and restricting which methods can carry request bodies. ### How HTTP Terminator Works **Kettle** detailed in a technical write-up that **HTTP Terminator** was fed 138 HTTP and SMTP RFCs, fragmented into roughly 15,000 small pieces. These fragments served as inspiration for generating the 30,000 unique candidate vectors. The system leveraged models like **Claude** for document extraction and test-case generation, with its investigator stage requiring **Claude Code**. One notable discovery involved a `Content-Type: multipart/byteranges` technique that successfully bypassed multiple server implementations, exposing over 200 websites in the test set, including a major U.S. bank. ### Open-Sourcing the Research Tool **PortSwigger** has open-sourced **HTTP Terminator** on GitHub, making the tool available to the broader security community. This move encourages further research and development in AI-assisted vulnerability discovery. ### Broader Research in Desync Attacks Separately, researchers behind CRLF-powered desync attacks have released public tools like `crlf-desyncs` and `crlf-powered-desync-scanner` to aid in studying and detecting this class of attacks. **Kettle** also reported that newer models, such as **GPT-5.6 Sol**, achieved a 30% success rate on a rediscovery benchmark when provided with an inspiration technique, indicating the growing potential of AI in security research.