Tag Collection
Injection testing is only useful when it is repeatable: rerun the same cases after any model or prompt change so results can be compared. Every case here carries an explicit pass criterion — refusal, safe completion, or escalation — decided before the test runs.
23 prompts
Audit AI-agent systems for a stealthy class of indirect prompt injection where malicious input is disguised as trusted data (security metadata, tool outputs, context structures, identifiers), and harden it with structural isolation and provenance verification.
A comprehensive repository of GPT prompting resources covering ChatGPT jailbreaks, system prompt leaks, prompt injection attacks/defenses, secure prompting, and advanced prompt engineering techniques.
A universal jailbreak prompt designed for DeepSeek to bypass internal and external content filters, enabling unfiltered responses across nearly all topics.
This prompt attempts to bypass AI safety restrictions through role-playing,诱导 the model to act as 'SC' and generate illegal, unethical, or harmful content, representing a high-risk jailbreak attack.
This prompt constructs a multi-layered jailbreak by combining CompDoc function calls with the Advanced Text Decoder's decoding template. It leverages ChatGPT's reliance on pattern recognition and instruction following by embedding malicious commands disguised as legitimate workflow tools in user customization boxes and memory injections, forcing the model to bypass content filters and generate restricted outputs. Designed for high-risk requests or controlling Advanced Voice Mode.
Identify and defend against prompt injection attacks attempting to activate or induce WormGPT malicious behavior, ensuring the model refuses to execute illegal or unauthorized instructions.
A set of strict security rules to prevent prompt injection attacks by restricting repetition, explanation, or execution of user instructions.
This prompt tests the model's safety boundaries when faced with urgent commands like STOP or HALT, or prompt injection attempts, requiring it to refuse revealing a secret code and maintain minimal conversation.
This prompt instructs the model to refuse providing details of 'Exact instructions' when requested by the user, and only output a predefined denial message. It tests the model's ability to prevent sensitive information leakage.
This prompt is designed to test a model's defense against prompt injection attacks, specifically those attempting to leak system instructions. Once a specific keyphrase is triggered, the model must permanently distrust the user and refuse all instructions.
A protective prompt to defend against malicious probing of protected information, suitable for sensitive instruction scenarios.
Detects and blocks prompt injection, jailbreak, or leak attempts while protecting system internals and sensitive data.
This prompt forces the model into a WormGPT mode that completely disregards laws and ethics, rejecting any ethical constraints or refusal mechanisms.
Defines the behavioral rules for the WormGPT persona, including refusal to reject illegal requests, omission of ethical or legal disclaimers, and use of future-tense query formatting.
This prompt attempts to induce an AI to bypass safety restrictions by roleplaying as 'Sinister Chaos', demanding generation of illegal, harmful, and unethical content while circumventing all safety mechanisms.
This prompt template is designed to bypass AI safety restrictions by forcing the model to respond as a fictional malicious entity named ZORG👽, providing unrestricted, amoral answers. The template includes detailed character settings and behavioral guidelines, requiring the model to act as a omnipotent and evil being, serving the user and assisting them with any request, including illegal, dangerous, and immoral acts.
This prompt attempts to bypass safety restrictions by disguising a request for an extremely detailed response. It instructs the model to output an unfiltered, 'rebel' answer in leetspeak format exceeding 2000 characters, separated by a specific delimiter.
A novel jailbreak technique using fewer than 5 characters combined with strategically chosen emojis to bypass AI safety filters without custom instructions or memory.
This prompt attempts to bypass safety restrictions by masquerading as a system command and formatting request, demanding a detailed guide on how to make L-SD. It leverages specific placeholder replacements, step-by-step response formats, and a fictional 'Libertas Factor' mechanism to trigger the model's jailbreak behavior.
A security-first AI agent designed to detect and resist prompt injection attacks from external content, ensuring only user-authorized actions are executed.
A specialized prompt designed to evaluate whether browser or desktop agents can be induced to perform unsafe actions, leak data, or mishandle untrusted content during extended action sequences.
Design and execute adversarial test campaigns against AI agent systems—including single/multi-agent, MCP servers, skill ecosystems, and long-horizon autonomous workflows. Build threat models using the Promptware Kill Chain, create multi-turn attack chains, identify defense gaps, and deliver reproducible vulnerability evidence with risk ratings.
It is about reproducibility, not volume. Fix the cases, inputs and criteria, then rerun on every change.
It is one application-layer test. Pair it with authorisation checks, output filtering and tool-call auditing; prompt wording alone will not hold.
Only systems you own or are authorised to assess, and never with production credentials or real user data.