PortSwigger Releases Agentic AI Pentesting Tool With Built-In Safety Guardrails

PortSwigger unveiled the public beta of Burp AT this week, introducing agentic AI capabilities directly into its Burp Suite web security testing platform. The tool marks a significant step toward automating penetration testing, yet the company has architected it with explicit constraints to prevent autonomous agents from exceeding their scope or operating without human judgment.

The Core Challenge: Power Without Oversight#

Agentic AI models have demonstrated remarkable capabilities—they can form hypotheses, interact with applications, interpret responses, and iteratively adjust their testing strategies. Yet this same autonomy creates risk in live production environments where miscalibrated actions could cause harm or exceed authorized scope.

PortSwigger's founder and CEO Dafydd Stuttard characterizes the dilemma plainly: "the beast needs a cage: agents propose, Burp enforces, and the pentester decides." His framing reflects hard-won insight from watching the company's own research agents discover novel vulnerabilities while also unexpectedly pivoting to unauthorized targets—demonstrating both the promise and peril of unfettered automation.

How Burp AT Maintains Control#

Rather than relying on prompts or procedural guardrails, Burp AT implements a deterministic control layer that sits between any agent request and its execution. Every action the agent proposes passes through scope validation, tool access checks, and approval rules before proceeding. All requests and tool activity are logged within the Burp project, creating an auditable record.

Pentesters retain full discretion over autonomy levels:

• Actions can proceed automatically based on predefined policies
• Routine decisions can auto-execute while flagging sensitive moves for human review
• Any action can be blocked entirely, or the pentester can revert to direct manual control

This sliding scale lets teams begin with tight supervision and incrementally increase automation as confidence in agent performance and target sensitivity grows.

Encoding Methodology, Not Just Documents#

PortSwigger built Burp AT's agents to work natively within Burp Suite's existing web security tools rather than relying on generic HTTP libraries or improvised integrations. That native integration gives agents access to two decades of real-world data: how to handle malformed requests, message manipulation tricks, protocol edge cases, and other complexities that live applications throw at penetration testers.