Deconstructing the Architecture of AI-Orchestrated Web Attacks

The landscape of web security has reached a critical inflection point, shifting from a battle of human keystrokes to an era of industrialized exploitation. Our research in the Akamai Security Intelligence Group (SIG) revealed a significant compression of traditional security remediation timelines. Although security teams have historically relied on multiday patch cycles to fix known vulnerabilities, automated machine-scale attacks now require immediate, real-time defenses. 

We call it the “speed gap” — the mathematical impossibility that human-led defense can keep pace with machine-scale offense. DARPA Challenge findings have demonstrated that AI systems can identify dozens of vulnerabilities in hours and generate functional exploits in less than 10 minutes. 

A persistent technical misunderstanding among security stakeholders is that frontier LLMs perform attacks directly. In fact, they orchestrate automated systems that execute vulnerabilities rapidly at scale. And human-led remediation can still stretch into days or months … or longer. So, security teams face a clear operational imbalance. 

This is the architecture of a “decoupled orchestration” model. The LLM acts as the orchestrator (or brain) by developing the exploit, performing deep analysis of source code, identifying out-of-band (OOB) testing opportunities, and chain-linking logic flaws.

The executor (or hands) of the operation are the secondary tools — headless browsers, custom Python scripts, or existing frameworks — that execute the actual web requests. 

The LLM operates behind the attacker's API, decoupled from the TCP/TLS handshake, while the hands execute the web requests. 

The table outlines how attackers leverage AI across each phase of the attack lifecycle, alongside the corresponding technical defenses we use to detect and mitigate these advanced threats.

^{Breakdown of AI-driven attack phases, corresponding automated techniques, and Akamai’s adaptive technical defenses}

By pairing autonomous threat logic with adaptive edge security, organizations can effectively contain and block sophisticated automated attacks before they cause severe impact.

The bridge between high-level reasoning and offensive execution is the Model Context Protocol (MCP). This protocol allows terminal-based AI assistants to drive security suites autonomously. A prime example is the Burp AI Agent, which integrates MCP tooling directly into the pen testing workflow.

By using automated capabilities like burp-scan, an AI assistant can independently operate security scanners. It uses a reasoning engine to evaluate 62 vulnerability classes — including injection, authentication, and cryptography — against a library of more than 200 payloads. Crucially for researchers, these tools now include sophisticated Privacy Modes (STRICT/BALANCED/OFF) to redact sensitive data, alongside passive analysis protocols that analyze traffic without sending a single probe.

From a WAF perspective, identifying these attacks is an exercise in detecting shadows. In offensive AI scenarios, there is no “User-Agent: LLM” because the LLM is never the source sending the actual web traffic. 

Because of the decoupled orchestration model, the security provider sees the result of the AI’s intelligence (the web request) rather than the AI itself.

Offensive AI orchestrators can use secondary tools to spoof their identity — meaning traditional, static indicators are useless (Figure). The abstraction through headless browsers ensures that the WAF sees a standard-looking browser session, while the “brain” sits miles away, refining the attack logic in real-time based on the feedback it receives from the “hands.”

To counter machine-scale attacks, we must deploy machine-scale defense. Relying on human intervention in a 10-minute exploit window is a recipe for failure. At Akamai, we leverage behavioral telemetry from across our global edge platform to identify machine-speed probing that eludes traditional filters.

We use the Akamai Inference Cloud and our distributed edge to close the speed gap via three strategies:

1. Identifying critical fault lines: We use behavioral analytics to detect the systematic mapping of API logic, identifying the “probing” phase before a weaponized payload is ever delivered.

2. Preemptively defending AI speed: Our threat research teams use AI-driven insights to surface potential bypasses, updating WAF rules globally before an AI-orchestrated attack can find a foothold.

3. Moving defense to the edge: By moving defense to the edge, we provide organizations with the ability to shield infrastructure instantly, ensuring that an AI-discovered bug tonight does not become a successful breach tomorrow.

The transition from human keystrokes to machine scale is complete. In today’s threat landscape, traditional reactive security is now a mathematically impossible winning strategy. When an orchestrator can iterate through thousands of evasive tactics in minutes, the defense must be equally autonomous and resilient.

The future of web security lies at the machine-scale edge. Organizations must move beyond the “patch and pray” model and adopt a proactive security posture that operates at the same velocity as the attackers. To secure the future of the modern enterprise, we must meet offensive AI at the edge.