The terms, in plain language.
Modern application and AI security testing has a vocabulary that often hides simple ideas. Here are the ones that matter, defined plainly — with a link to how we test each where there is more to say.
Broken access control
A vulnerability where an application lets an identity perform an action or reach data it should not. The failure is the gap between what the app returned and what its own authorization rules require — which is why it is invisible to a tool that looks at one request. It is the most common serious web risk. How we test it →
IDOR (Insecure Direct Object Reference)
A form of broken access control where a user reaches an object that is not theirs by changing an identifier — /invoice/1042 → /invoice/1043 — because the server does not verify ownership.
BOLA (Broken Object Level Authorization)
The API framing of IDOR, and number one on the OWASP API Security Top 10: an endpoint returns an object without checking that the caller is allowed to access that specific object. APIs made it worse because object identifiers are everywhere and the client is often the only thing enforcing "your own data only."
Prompt injection
An attack on an AI system where crafted input causes the model to follow attacker instructions instead of the intended task. The AI-era analogue of injection attacks, now number one on the OWASP Top 10 for LLM applications.
Second-order (indirect) prompt injection
Prompt injection where the attacker never talks to the model. They plant instructions in data the agent reads later during a normal task — a ticket, a document, a web page — so a legitimate user's session carries out the attacker's intent. The dangerous variant, and the one worth proving with a canary. How we prove it →
MCP (Model Context Protocol)
A standard way for an AI agent to connect to external tools and data sources. Useful, and an attack surface: each connected MCP server's tools can be abused, its data can carry injected instructions, and its permissions can be broader than the task needs. Testing AI agents and MCP →
DAST (Dynamic Application Security Testing)
Automated scanning of a running application by matching patterns on requests and responses. Fast, cheap and good at the known and obvious — CVEs, missing headers, misconfiguration — but structurally blind to access control, business logic and AI behaviour. Scanner vs. AI pentest →
PTaaS (Penetration Testing as a Service)
Delivering penetration tests through a platform and subscription rather than one-off engagements, often blending automation with human testers. SuperPentest is an autonomous take on the idea: continuous, re-runnable, with every finding reproduced.
Business logic vulnerability
A flaw in how an application's intended workflow can be abused — negative quantities, skipped steps, race conditions — that is only wrong relative to what the app is for. No generic signature detects it, because the rule being broken is specific to your application.
False positive
A reported issue that is not actually exploitable. Scanners produce many because they cannot confirm impact and leave triage to you. Reproduction-based testing holds a candidate back until it is proven, so what you receive is signal, not a queue.
Full definitions
Each term has a deeper page — a plain-language definition, a concrete example, and how SuperPentest tests for it:
Want the test behind a term? Read the research, a worked finding, or book a demo.