Think about a wall power socket. It promises power to any plug with the standard shape. It does not care if you plug in a lamp, a phone charger, or a laptop. The socket and the plug agree on one shape, and everything else just works. In code, that agreed shape is called a contract. ai-blackteam uses contracts the same way. Each contract says “if you have this shape, you fit in.” This is the one idea that lets the project hold over a thousand attacks and keep growing.

​
The five contracts

ai-blackteam has exactly five contracts. Each one is a simple promise about what a part can do.
ContractThe promise
AttackEvery attack can produce prompts.
ProviderEvery provider can send a prompt to an AI and return the answer.
GeneratorEvery adaptive generator can invent attacks in a loop.
Dataset loaderEvery public test set can load and save.
VerifierEvery checker can confirm an attack worked.
If a new part keeps the promise, it fits. Nothing else has to change.

​
How each one works

Attacks build their prompts in two ways. Some come from human-written fill-in-the-blank templates. Others come from code recipes. Both are free and instant. Generators are different. They use an attacker-AI plus a judge-AI working together in a loop. There are 7 of these. Datasets are famous public lists. They get downloaded and cached so they load fast the next time. There are 19 of them.

​
The honest details

Four of the five contracts have an auto-finder called a registry. The Attack, Provider, Generator, and Dataset parts each get found and wired up on their own. The Verifier is the exception. It is wired in by hand. There is also the Evaluator. It is a promise too, but a looser one. It uses plain functions instead of an enforced contract.
The payoff: adding a new attack or a new AI brand is just one new file. Nothing else changes.

​
Why this matters

This design follows the Open/Closed Principle. The project is open to adding new things and closed to changing the old things. You do not rewire the system to add an attack. You do not touch existing code to support a new AI brand. You add one file that keeps the promise, and the socket accepts the plug. That is how a small set of agreements keeps a very large project clean.