Insights
Autonomous AI Pentesting vs Human-Led VAPT: What Your RFP Should Still Require
Why This Question Is on the Table Now
Through 2026, autonomous and AI-driven penetration testing has moved from conference demos into products that security teams can buy. Several established security vendors and a number of start-ups now sell agents that discover assets, pick attack techniques, attempt exploitation and write up findings with little or no human steering. The pitch to a CISO is straightforward: test every week instead of once a year, at a fraction of the cost of a consultant-led engagement.
That pitch is landing in procurement. We are seeing RFPs where an autonomous platform is offered as a like-for-like replacement for a human-led VAPT. Sometimes that is fine. Often it is not, and the difference only shows up after a breach or an audit. This guide sets out where each approach is strong, where it falls short, and what an RFP should require so you get the benefit of automation without losing the assurance you are actually buying.
Where Autonomous Testing Genuinely Helps
- Frequency. An agent can retest the same estate every week or after every change. A yearly manual test is a snapshot; most environments drift within weeks.
- Breadth on known vulnerability classes. Missing patches, exposed services, default credentials, weak TLS, common misconfigurations and well-documented web flaws are exactly what automation finds quickly and consistently.
- Attack-path validation on internal networks. Chaining a foothold through credential reuse and privilege escalation in Active Directory is repetitive work that agents handle well, and showing a real path to a crown-jewel system is more persuasive than a list of isolated findings.
- Fix verification. Confirming that a remediated finding stays fixed is a good fit for continuous automated retesting.
- Pairing with attack surface discovery. Combined with external attack surface management, automation can test new internet-facing assets soon after they appear, instead of waiting for the next annual scope.
Where a Human Tester Still Finds What the Agent Misses
- Business logic. Paying less than the listed price, approving your own request, skipping a KYC step or reusing a one-time coupon are flaws in how the application is meant to work. Spotting them needs an understanding of the business process, not just the HTTP traffic.
- Authorisation across roles and tenants. Testing whether a branch user can see another branch's customers, or one tenant can reach another tenant's data, means building a model of who should see what. This is where many of the most serious findings in Indian BFSI and SaaS applications sit.
- Novel and chained issues. Combining three low-severity findings into one critical one, or finding a bug class nobody has documented yet, is still largely human work.
- AI and LLM applications. Prompt injection, data leakage through retrieval and over-privileged agent tools need judgement about what counts as harmful in your context. See our guides to LLM application penetration testing and AI and LLM clauses in a VAPT RFP.
- Severity in context. A tool rates a finding by its generic score. A tester rates it by what it exposes in your business, which is what your board, your regulator and your remediation queue actually need.
- Accountability. Regulated entities and their auditors generally expect a report signed by a named, qualified tester who can explain and defend every finding. A dashboard export does not do that on its own.
What Your RFP Should Require
Whether the bidder offers a human-led engagement, an autonomous platform or a mix, these clauses keep the comparison honest:
- Disclose the tooling and the level of autonomy. Ask which parts of the work are automated, which tools or platforms are used, and which steps a human performs. A bid that cannot answer this clearly is not comparable with one that can.
- Human validation of every reported finding. Each finding in the final report should be reproduced and confirmed by a named tester, with false positives removed before the report reaches you.
- Minimum manual coverage. State the areas that must be tested by hand: business logic, authorisation across every role, payment and approval flows, and any AI or LLM features. Ask for the test cases to be listed in the report.
- Guardrails for autonomous agents. Hard scope enforcement so the agent cannot wander outside agreed IP ranges and URLs, no destructive or denial-of-service techniques, rate limits, agreed testing windows for production, and a kill switch your team can trigger.
- Data handling. Autonomous tools may send requests, responses and screenshots to a cloud platform, and some route them through third-party language models. Ask where that data is stored and processed, who can access it, how long it is kept and when it is deleted. If the systems hold personal data, this is part of your own DPDP security obligations as the Data Fiduciary.
- Reproducible evidence. Every finding should come with the steps, requests and evidence needed for your developers to reproduce it, not just a title and a score.
- Retest and sign-off. Include a retest of critical and high findings and a final report that shows closure, signed by the tester responsible.
A Practical Model: Continuous Automation, Periodic Human Depth
For most organisations the answer is not either-or. Use autonomous testing for breadth and frequency between engagements, so drift and newly exposed assets are caught in days rather than months. Then commission a human-led VAPT at least once a year, after major releases and before compliance milestones, focused on business logic, authorisation, AI features and the attack paths the automation surfaced. The automated results make the human engagement better, because the tester starts from a cleaner baseline and spends time on the hard problems instead of the obvious ones.
If you are regulated, check what your sector rules expect before relying on automation for a mandated assessment. Our SEBI CSCRF guide is a starting point for capital-markets entities.
Common Mistakes We See
- Comparing price per test without comparing coverage. A cheap weekly scan and a manual application test are different products. Compare what each actually tests.
- Running agents in production with no rules of engagement. An autonomous agent that locks accounts or floods a payment API during business hours causes a real incident.
- Treating "no critical findings" as clean. Automation reports what it knows how to find. Absence of findings in business logic usually means the area was not tested.
- Forgetting where the test data goes. Screenshots of customer records sitting in a vendor's cloud are a data exposure of their own.
eNeoteric runs human-led VAPT engagements and security services, and can review an RFP or a bid that proposes autonomous testing, so you know what is and is not covered. Email [email protected] or talk to our team.
Explore all ← Back to Insights