Agentic AI in web testing is changing how vulnerabilities are found

At a glance

What this is: PortSwigger highlights how agentic capabilities in Burp AI are being used in Repeater to support discovery of web vulnerabilities such as SQL injection, XSS and insecure configurations.

Why it matters: This matters because AppSec teams are starting to fold AI-assisted reasoning into testing workflows, which changes how they validate findings, prioritise reviews and govern human oversight.

Context

Agentic AI in web testing refers to AI-assisted workflows that can reason over targets, guide test steps and help surface likely weaknesses without replacing the tester. The governance gap is that faster discovery can look like stronger assurance even when the underlying control coverage has not changed.

This article sits at the intersection of application security and AI-assisted operations, with a genuine identity angle only in the way human testers supervise tool use. That means the core question for practitioners is not whether the AI can help, but how to keep evidence, reproducibility and accountability intact when testing becomes more automated.

Key questions

Q: How should security teams govern AI-assisted web testing tools?

A: Treat AI-assisted testing as a governed workflow, not a convenience feature. Define which targets, data, and actions the tool may touch, assign separate credentials and logs, and require human approval for anything that could affect production systems. The goal is to keep the agent’s scope narrow enough that its actions remain attributable, reviewable, and reversible.

Q: Why can agentic testing speed up discovery without improving assurance?

A: Agentic testing can surface more candidate issues faster because it accelerates search and reasoning, but assurance still depends on reproducible evidence, contextual validation and clear scoping. Faster output can increase coverage, yet it does not by itself prove exploitability or business impact.

Q: What are the signs that an AI pentesting system is producing unreliable findings?

A: Common warning signs include findings that lack concrete exploitation evidence, reports based only on weak surface indicators, and inconsistent results when the same test is repeated. A high false positive rate is another signal. If the system often infers a vulnerability from suspicious behavior without proving impact, its conclusions need tighter verification.

Q: What should AppSec teams do when AI is used in bug bounty style testing?

A: Teams should set explicit rules for scope, supervision and evidence handling before using AI in bug bounty style testing. That means documenting the prompts used, confirming that humans still own the final judgment, and ensuring the workflow can withstand review if a result becomes a security ticket.

Technical breakdown

How agentic AI changes the testing workflow

Agentic capabilities in a web testing tool shift some of the work from manual pattern search to guided reasoning over requests, parameters and responses. In practice, the model can propose likely attack paths, adapt prompts and interpret outputs from Repeater, which is useful when exploring SQL injection, cross-site scripting and configuration flaws. That does not make the system autonomous in a security sense. The tester still defines scope, validates exploitability and decides whether a finding is real. The mechanism matters because the AI is accelerating analyst judgment, not replacing the evidentiary standard for a vulnerability claim.

Practical implication: Treat agentic output as investigative support, not as a proof of compromise.

Why prompt design now affects finding quality

Prompting becomes part of the testing method because the model’s search behaviour is shaped by the instructions it receives. A vague prompt can produce broad, noisy suggestions, while a well-formed prompt can focus the model on a specific parameter, input context or vulnerability class. In AppSec terms, that creates a new control surface around test design, reproducibility and auditability. If two testers use different prompts, they may not reproduce the same result even against the same target. The practical issue is not just whether the model is capable, but whether the workflow can be repeated and defended under review.

Practical implication: Standardise prompt patterns and preserve the exact test inputs alongside each finding.

Where AI-assisted pentesting still needs human validation

Web testing still depends on distinguishing suspicious behaviour from exploitable weakness. AI can help identify anomaly patterns, but it cannot on its own prove impact, confirm business logic abuse or establish that a result is not a false positive. That is especially important for insecure configuration findings, where context determines severity, and for injection testing, where a suspected issue may fail under controlled verification. The control gap is evidentiary, not computational: if the team cannot explain why a result is valid, the finding is not ready for remediation prioritisation. Human review remains the accountability layer.

Practical implication: Require manual confirmation before a finding enters remediation or board reporting.

NHI Mgmt Group analysis

Agentic AI changes the economics of AppSec before it changes the governance model. The immediate effect is faster discovery of candidate issues, not a fundamentally different security objective. That means teams will feel pressure to widen testing coverage while keeping the same review standards, and that tension will expose weak evidence handling. Practitioners should treat AI-assisted testing as a throughput change, not an assurance shortcut.

The real control question is reproducibility of findings, not whether the model can find flaws. If the prompt, context or session state changes the result, the workflow becomes hard to audit and hard to defend in an incident review. In that sense, agentic testing introduces a documentation problem as much as a detection problem. Practitioners need traceable inputs, outputs and reviewer sign-off.

AppSec programmes will need clearer boundaries between exploration and validation. An AI system may be useful for broad reconnaissance inside a test scope, but the evidentiary threshold for a vulnerability report still belongs to the human tester. This is where governance becomes operational: teams must define when AI can suggest, when it can prioritise and when it must stop. Practitioners should codify those handoffs now.

AI-assisted test governance: the new risk is not that AI finds too much, but that teams cannot consistently explain what the AI actually did. That creates a traceability gap across prompts, test paths and reviewer decisions. AppSec leaders should build evidentiary controls around AI-assisted workflows before they rely on them for scale.

What this signals

The practical shift for AppSec teams is that AI-assisted testing now sits inside the control plane for vulnerability discovery, which means governance has to move closer to the test session itself. If the team cannot reconstruct what the model was asked to do, the evidence trail is too weak for dependable triage.

AI-assisted evidence trail: the important governance task is to capture prompts, request context and reviewer sign-off as part of the finding lifecycle. That turns AI from an opaque helper into an accountable testing component. Practitioners should design for auditability before they scale use of agentic workflows.

For practitioners

• Define AI-assisted testing boundaries Specify which stages of web testing the AI may support, such as candidate discovery, request crafting or result summarisation, and require human validation before any issue is recorded as a finding.
• Preserve prompt and session evidence Store the exact prompts, request context and output history for every AI-assisted test so reviewers can reproduce the path that led to the finding.
• Separate exploration from verification Use AI to broaden search coverage, then require a deterministic manual check for exploitability, business impact and false-positive filtering.
• Update AppSec review criteria Add review steps for AI-assisted findings that check whether the issue was reproducible, scoped correctly and supported by sufficient evidence for remediation priority.

Key takeaways

• Agentic AI in web testing can accelerate discovery of SQL injection, XSS and configuration weaknesses, but the underlying assurance model still depends on human validation.
• The key governance problem is not model capability alone, but whether AI-assisted findings can be reproduced, explained and audited.
• AppSec teams should codify where AI may assist, where humans must verify, and what evidence must be retained for every finding.

Key terms

• Agentic Testing: Security testing in which software agents can choose actions, sequence steps, and adapt during a validation workflow. In practice, it combines automation with governance, because the agent is not just running a script. It is operating with enough decision-making to require scope limits, auditability, and oversight.
• Reproducibility: Reproducibility is the ability to recreate a model’s output from the same or equivalent inputs and configuration. In governance terms, it is a proof that the training process is sufficiently recorded to support validation, audit, and rollback when outcomes need to be challenged.
• Evidence Trail: An evidence trail is the set of records that explains how an identity decision was made, including inputs, checks, outcomes, and escalations. It matters because regulated onboarding must be defensible after the fact, not just successful in the moment.
• Exploratory Discovery: Exploratory discovery is the broad search phase of testing where the goal is to surface candidate weaknesses rather than prove exploitation. AI can be useful here because it can accelerate search, but the output still requires deterministic validation before it becomes a remediation-grade finding.

Deepen your knowledge

NHI Mgmt Group covers identity security, NHI governance, and agentic AI through independent research, practitioner guides, and the NHI Foundation Level course. Explore the course when you need a structured way to connect governance, identity and operational control across modern security programmes.