The publication presents a collection of prompt injection vectors and strategies, accompanied by examples. The material may help engineers recognize such techniques while evaluating systems.
When studying the examples, treat instructions found in prompts as untrusted input. Consider what behavior each technique attempts to trigger, and record observations without running suspicious content in production systems.
Use the cases to structure controlled evaluations: define expected behavior, test examples in an appropriate environment, and document the results. The collection can support risk identification, but it does not guarantee that a system is protected.
If you use AI to analyze or adapt the material, avoid submitting personal data, credentials, secrets, or internal information. Prefer synthetic examples and check the chosen tool’s retention and access policies.