HTTP Terminator Proves AI Can Autonomously Discover Attack…

James Kettle demonstrates that PortSwigger's HTTP Terminator AI system independently generates HTTP desynchronization techniques. The Apache Traffic Server zero-day tracked as CVE-2026-63078 remains unverifiable in public databases.

On August 7, 2026, James Kettle, director of research at PortSwigger, published the results of HTTP Terminator, an AI-assisted research system designed to generate and validate HTTP desynchronization vectors at industrial scale. The system autonomously produced previously unknown attack techniques, tested against roughly 30,000 authorized websites, with immediate consequences for critical sectors including banks and government infrastructure. The research surfaces a concrete operational problem: the zero-day identified in Apache Traffic Server, tracked as CVE-2026-63078, is not verifiable in public databases as of publication.

Key Takeaways

HTTP Terminator autonomously generated and demonstrated new HTTP desynchronization techniques without direct human discovery input, according to the cited source.
The system tested approximately 30,000 candidate vectors against roughly 30,000 authorized sites, identifying approximately 700 vulnerable targets before deeper validation.
A technique based on Content-Type: multipart/byteranges exposed over 200 sites in the test set, including an unnamed U.S. bank.
The Shared-Parser Confusion concept was proposed by the AI and validated by Kettle, while the Apache zero-day CVE-2026-63078 required a human-guided discovery cascade.

How Autonomous HTTP Research Works

HTTP Terminator operates by ingesting standard technical documentation and transforming it into inspiration for attack vectors. According to the cited source, Kettle fed the system 138 HTTP and SMTP RFCs, split into roughly 15,000 fragments used as seeds for generation. From this input the system produced approximately 30,000 candidate desync vectors, then tested them against a pool of roughly 30,000 websites where scanning was authorized via bug bounty or vulnerability disclosure programs.

The source reports that roughly 700 vulnerable targets surfaced before deeper validation and further Request Queue Poisoning (RQP) research. Targets included banks, government infrastructure, security products, and an airport, though specific names were not disclosed. This sectoral distribution indicates the attack surface intercepted by the system is not limited to marginal endpoints.

The Line Between AI Autonomy and Human Intervention

The research draws a sharp line between what the AI generated on its own and what required human validation. According to the cited source, HTTP Terminator produced and demonstrated several techniques "without direct human discovery input." However, the path to the Apache zero-day and the Shared-Parser Confusion concept followed different dynamics.

The dangling-byte technique represents a purely autonomous discovery that was validated. The source reports the system tested 16 ideas to improve RQP and only this technique survived evaluation: it leaves a smuggled request one byte short, eliminating a race condition that makes RQP unreliable. By contrast, Shared-Parser Confusion — which emerges when response processing rules are mistakenly applied to requests because servers reuse parsing logic — was "proposed by the system but validated by Kettle," as the researcher's direct quote states: "Neither of us would have discovered it alone."

"Neither of us would have discovered it alone" — James Kettle, on Shared-Parser Confusion

The CVE-2026-63078 Case: An Operational Verifiability Gap

A human-guided discovery cascade exposed a vulnerability in Apache Traffic Server, tracked as CVE-2026-63078. As of August 7, 2026, however, direct verification in public databases failed: the CVE does not appear in CVE.org or NVD, and the Apache advisory from July 2026 listing 34 flaws did not include it, according to the source.

This information void is not marginal. For defensive teams, the absence of an official record means no guidance on patched versions, CVSS severity, or concrete mitigations. The source does not specify which Apache Traffic Server version resolves the issue, nor a patch release date. The CVE identifier exists in the research narrative but not in the governance of reference databases, creating a risk condition that cannot be quantified through standard channels.

The Released Implementation and Models Involved

PortSwigger has open-sourced HTTP Terminator. The released implementation uses Claude for document extraction and test case generation, while the investigator stage requires Claude Code, according to the cited source. The system does not rely exclusively on a single model: Kettle tested newer models on a rediscovery benchmark, reporting a 30% success rate for GPT-5.6 Sol when equipped with an inspiration technique.

The decision to make the tool public raises accessibility questions. Democratizing HTTP desync research — a vulnerability category notoriously difficult to detect with conventional tools — can accelerate defense but also lower the barrier for attackers with access to the same AI infrastructure.

Immediate Actions

Verify Apache Traffic Server adoption in your infrastructure and monitor for the emergence of CVE-2026-63078 in official databases (NVD, CVE.org, Apache advisory), given it is not currently traceable.
Evaluate transitioning away from HTTP/1.1 upstream where architecturally feasible, as the source reports this remains the primary recommended defense.
Where HTTP/1.1 upstream cannot be removed, implement method allow-listing at both levels of the proxy chain and restrict methods that can carry a body, as indicated by the cited source.
Cautiously examine adoption of HTTP Terminator or analogous tools in authorized research contexts, considering the open-source code is available but its completeness and stability are not independently verifiable from available sources.

Why Verifiability Can't Keep Pace with Discovery

Kettle's experiment exposes a structural tension: vulnerability disclosure and cataloging systems are designed for human-speed research, while AI can generate and test vectors at industrial scale in compressed timeframes. CVE-2026-63078 exists in the discovery narrative but not in the databases defenders consult daily. This gap is not attributable to a procedural error; it is the consequence of technical acceleration outpacing governance infrastructure.

The source does not document in-the-wild exploits for the discovered techniques. However, the presence of banks and government infrastructure in the vulnerable target set indicates the impact surface is critical. Open-sourcing HTTP Terminator makes this discovery capability replicable: the line between defensive and offensive research will depend increasingly on authorization context, not technical availability.

The research suggests the most productive human-AI collaboration model is not replacement but intersection. Pure desync techniques emerged from system autonomy; deeper concepts like Shared-Parser Confusion emerged from the overlap between algorithmic proposal and expert validation. The operational problem is that defenders still lack tools to verify at the same velocity.

Sources

Information verified against cited sources and current as of publication.

Sources

Sources and references