A new AI system called HTTP Terminator has identified hundreds of websites vulnerable to HTTP request smuggling attacks, successfully breached them at scale, and discovered what researchers are calling a "genuinely new class" of vulnerability known as "shared-parser confusion." But the finding that may matter most is this: the AI didn't work alone—it was guided by a human researcher throughout the entire process, a collaboration that challenges the prevailing narrative about autonomous AI research, according to a white paper published by PortSwigger, a security company whose director of research built the system.
HTTP Terminator created 30,000 distinct attack vectors and tested them continuously, soon flagging 700 vulnerable targets that included financial institutions and government infrastructure, according to the research. The system successfully extracted a live API key from one bank during testing. Multiple breaches were linked to weaknesses in widely used enterprise products including Apache Traffic Server, Citrix NetScaler, and BeyondTrust. The AI was able to invent and validate several novel desync triggers, one novel desync pattern, and a desync weaponization technique—all through live testing on websites that permit security research via bug-bounty programs and Vulnerability Disclosure Programs.
"This inverts the accepted narrative by showing an expert can be a massive amplifier for an AI research system," James Kettle, PortSwigger's director of research, wrote in the paper. "A human in the loop can still add significant value, as opposed to just building the loop, then stepping back." Kettle emphasized that the shared-parser confusion discovery, which he called "absolutely huge," wasn't fully autonomous: "the HTTP Terminator proposed it, and I validated it. Neither of us would have discovered it alone." The researcher used his own processes to design and construct the AI, posed narrow high-value questions, eliminated weak answers, applied anomaly-detection logic, used deterministic code to limit agent behavior, and applied discoveries to subsequent cascade research.
HTTP request smuggling attacks, also known as HTTP desync attacks, interfere with how websites process HTTP requests from multiple users. The vulnerability stems from what the report describes as inherently "weak" isolation of requests: front-end servers typically bundle several requests together over the same back-end network connection to boost performance and efficiency. When front-end and back-end systems disagree on where each request ends, attackers can send ambiguous messages that the back end interprets as two separate requests. Through response query poisoning, those attackers can inject false information into later requests or manipulate data flows so systems lose track of which responses correspond to which users, allowing them to intercept responses meant for others—sometimes containing sensitive information, credentials, or API keys. While these vulnerabilities are typically associated with HTTP/1 requests, sites supporting HTTP/2 may also be at risk depending on their back-end architecture, the research notes.
Kettle has open-sourced HTTP Terminator along with his research blueprint so other security teams can "turn their own methodology and instincts into an autonomous research weapon." He recommends tackling tasks in four steps: set an objective, create an evaluation strategy, establish inspiration sources, and explore cascade routes for new discoveries. The report advises asking high-value questions without being too broad, reviewing outputs in initial test runs to rule out low-value hypotheses, and remembering that "every extra sentence of prompt risks context-contamination." When a system is fully AI-driven and also relies on AI-built code, it's nearly impossible for it to improve over time, the research concludes. It's better to start quickly with an AI-heavy approach, then shift responsibility to deterministic code for speed and accuracy. "Fully autonomous research is real," Kettle acknowledged, but "humans are a massive power amplifier for AI research systems." The framework suggests that framing security research as "AI vs code vs human" is more productive than debating autonomous versus human-led approaches—because the highest-impact discoveries emerge when all three work together, each amplifying what the others can't achieve alone.