AI pentest glossary
In short
The key terms of AI-assisted penetration testing, defined plainly. Every definition has its own anchor so it can be cited or shared directly.
- Penetration test (pentest)
- An authorised security assessment in which an expert attempts to compromise a system as an attacker would, in order to identify and prove exploitable vulnerabilities.
- AI pentest
- A penetration test in which AI agents automate reconnaissance, analysis and drafting, under the control of a human tester who validates every vulnerability.
- Autonomous pentest
- A penetration test driven entirely by AI agents that choose their own actions. Fast, but prone to false positives and risky actions without human supervision.
- AI agent
- A program that observes a result, reasons with a language model, chooses an action (run a tool, send a request), then repeats based on what it observes.
- Multi-agent architecture
- A set-up in which several specialised AI agents (reconnaissance, analysis, verification, reporting) share an engagement and pass on their findings.
- Adversarial qualification
- A step in which one AI agent tries to disprove each vulnerability proposed by the others. Only findings that survive reach the human tester.
- False positive
- A reported vulnerability that does not actually exist. False positives waste remediation time and undermine a report's credibility.
- Hallucination
- A plausible-sounding statement by a language model that is not grounded in observed facts, such as a vulnerability declared exploitable without evidence.
- Pseudonymisation
- Replacing identifying data (IP addresses, URLs, names) with tokens such as url1 or ip1 before processing by a language model.
- Black, grey and white box
- The level of information given to the tester: none (black), user accounts (grey), or source code and documentation (white).
- Attack surface
- Every point through which an attacker can try to get in: websites, APIs, exposed services, remote access, user accounts.
- Vulnerability scanner
- A tool that compares a system with a database of known flaws. Useful continuously, it does not detect business logic flaws or attack chains.
- IDOR
- Insecure Direct Object Reference: an access control flaw that lets a user read or change another user's data by altering an identifier in a request.
- BOLA
- Broken Object Level Authorization: the API equivalent of IDOR, ranked first in the OWASP API Security Top 10.
- SQL injection
- A flaw that lets an attacker insert instructions into a database query through poorly filtered user input.
- Prompt injection
- An attack on an application that uses a language model: instructions hidden in an input hijack the model's behaviour. First category of the OWASP Top 10 for LLM Applications.
- CVSS
- Common Vulnerability Scoring System: a 0 to 10 severity rating for vulnerabilities. Version 4.0 was published in 2023.
- OWASP Top 10
- The reference ranking of the most critical security risks for web applications, published by the OWASP Foundation.
- PTES
- Penetration Testing Execution Standard: a framework that structures a penetration test into phases, from scoping to reporting.
- MITRE ATT&CK
- A knowledge base of real-world attacker tactics and techniques, used to describe and compare the actions taken during a test.
- Active Directory
- Microsoft's directory service managing accounts and permissions on a corporate network. Compromising it often means controlling the whole environment.
- Retest
- A follow-up test, performed after remediation, that verifies reported vulnerabilities have actually been fixed.
- NIS2
- EU Directive 2022/2555, which extends cybersecurity obligations to many more entities and requires them to assess the effectiveness of their security measures.
- DORA
- EU Regulation 2022/2554 on digital operational resilience for the financial sector, applicable since January 2025, which includes resilience testing requirements.