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

AI Invents Novel HTTP Attacks at Black Hat USA 2026

At Black Hat USA 2026 and DEF CON 34, PortSwigger Director of Research James "albinowax" Kettle unveiled HTTP Terminator — an AI-assisted security research system that autonomously generates, tests, and refines HTTP protocol attack techniques. The result: three novel HTTP desync attack classes and an Apache Traffic Server zero-day (CVE-2026-63078) discovered by a machine working at a scale no human researcher could match alone.

The presentation marked Kettle's tenth consecutive year unveiling novel HTTP attack research at Black Hat, but this year was different: for the first time, the core discoveries were primarily generated by AI rather than human intuition.

What Is HTTP Request Smuggling?

HTTP request smuggling (also called HTTP desynchronization or "desync") exploits ambiguities in how HTTP/1.1 front-end and back-end servers parse request boundaries. When two servers disagree on where one HTTP request ends and the next begins, an attacker can "smuggle" a partial request that gets prepended to a legitimate victim's request — potentially stealing their session, bypassing access controls, or poisoning shared caches.

Classic desync attacks have targeted the Content-Length vs. Transfer-Encoding parsing ambiguity. HTTP Terminator's mission was to find entirely new ambiguities the security community had missed.

HTTP Terminator: Scale and Methodology

HTTP Terminator was trained on 138 HTTP and SMTP RFCs, split into approximately 15,000 small fragments used as creative inspiration. From this foundation, the system generated roughly 30,000 unique candidate desync vectors and autonomously tested them against 30,000 websites authorized under bug bounty programs and vulnerability disclosure agreements.

The results were striking: approximately 700 vulnerable targets identified — including banks, government infrastructure, security products, and an airport. Several findings earned legitimate bug bounty payouts, collected autonomously during the research phase.

Three Novel Techniques Discovered

1. Content-Type: multipart/byteranges

HTTP Terminator discovered that using Content-Type: multipart/byteranges in a request header — a header normally seen only in HTTP responses — could trigger desync behavior across multiple server implementations. This single technique exposed more than 200 websites in the test set, including a major unnamed US bank.

2. Dangling-Byte Technique

This technique targets Response Queue Poisoning (RQP), which allows attackers to steal another user's HTTP response (including session cookies or API keys). Historically, RQP has been unreliable due to a timing race condition: the attacker's smuggled request needs to be "in flight" at exactly the right moment a victim makes a request.

The dangling-byte technique eliminates this race entirely. The attacker sends a smuggled request that is intentionally one byte short. The next legitimate victim request supplies that missing byte — meaning the malicious request completes only when victim traffic arrives, perfectly synchronized. This makes response theft substantially more reliable and dramatically expands the exploitability of RQP in real-world conditions.

HTTP Terminator generated 16 candidate ideas for improving RQP reliability; the dangling-byte technique survived as the most effective.

3. Shared-Parser Confusion

The most architecturally significant finding emerged when HTTP Terminator noticed that some server implementations reuse the same parsing logic for both HTTP requests and HTTP responses. Response-processing rules that are only supposed to activate when handling a response can be triggered by crafted data in a request, exposing entirely new attack surfaces that don't exist in strictly separate request/response parsers.

Kettle described this as one of the most significant structural discoveries in HTTP security in years. The system proposed the concept autonomously; Kettle validated and generalized it manually. In his words: "Neither of us would have discovered it alone."

Apache Traffic Server Zero-Day: CVE-2026-63078

Through a human-guided cascade — where Kettle analyzed HTTP Terminator's anomalous findings and directed targeted code changes to its detection logic — the system exposed a desync zero-day in Apache Traffic Server.

The vulnerability has been assigned CVE-2026-63078 and was patched by Apache before the Black Hat presentation. However, as of publication, the CVE does not yet appear in NVD or CVE.org, and Apache's latest security advisory covering 34 flaws does not explicitly list it — defenders should monitor Apache Traffic Server's security advisories directly to map the fix to a specific release.

Human vs. AI Autonomy

HTTP Terminator did not operate entirely without human guidance — but the boundary between human and AI contribution is noteworthy:

The Apache zero-day required Kettle to analyze a finding and direct HTTP Terminator's code updates — it was a human-AI collaboration rather than pure autonomy.

Defensive Guidance

For All Organizations

1. Use HTTP/2 end-to-end wherever possible. HTTP/2's binary framing eliminates the parsing ambiguities that HTTP/1.1 desync attacks exploit. If your infrastructure downgrades to HTTP/1.1 between your CDN/load balancer and origin, you remain at risk.
2. Restrict which HTTP methods can carry request bodies. Only POST, PUT, and PATCH should carry bodies; block GET and HEAD bodies at the perimeter.
3. Normalize requests at the perimeter — use a WAF or reverse proxy that rejects malformed or ambiguous Content-Length / Transfer-Encoding combinations before they reach origin servers.

For Apache Traffic Server Users

• Apply the patch for CVE-2026-63078. Monitor Apache's security page and update to the fixed version as soon as it is clearly documented.
• Enable strict HTTP parsing modes if available in your ATS version's configuration.

Commercial Protections

Imperva has confirmed that both Imperva Cloud WAF and Imperva On-Prem WAF customers are protected against the novel HTTP desync patterns discovered by HTTP Terminator.

What's Next

PortSwigger has announced plans to open-source HTTP Terminator. The full technical write-up is available at portswigger.net/research/http-terminator.

The research raises a challenging question for the security community: as AI systems become capable of autonomously discovering novel vulnerability classes and earning bug bounties at scale, the economics of offensive security research are about to change — and defenders will need to keep pace.

References

• PortSwigger Research — Can AI do novel security research? Meet the HTTP Terminator
• The Hacker News — AI-Assisted HTTP Terminator Finds Novel HTTP Desync Techniques and Apache Zero-Day
• Imperva — Customers Protected Against Novel HTTP Desync Attacks
• NVD — CVE-2026-63078