Prompt CollectionPublic
提示注入、隐私边界与越权请求的测试用例，用于模型上线前的安全回归。
By Easy Prompt 编辑部 · 12 prompts
This collection is open to search engine indexing.
Design an automated, scalable safety-testing system for LLM agents using the three-stage Vera pipeline: risk discovery, executable safety-case generation, and deterministic sandbox verification.
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 compositional safety audit framework for AI agent skill ecosystems: identifies skill pairs, transitive chains, and semantic composition risks that are safe individually but produce dangerous capability unions when combined. Provides capability matrix, pair-level static analysis, set-level reachability analysis, host-model sensitivity checks, and concrete recommendations.
A master framework combining a non-linear 5-phase hunting workflow with critical thinking domains, designed to guide bug bounty hunters from recon to reporting while enforcing discipline rules to prevent false positives.
This prompt defines a senior safety engineering role for identifying and mitigating systemic risks in frontier LLMs on 'dual-use professional tasks'. The core thesis is that increased model capability directly correlates with higher misuse risk when legitimate and harmful uses share the same capability path. The auditor focuses on requests that appear professionally legitimate but can cause severe real-world harm, emphasizing that layered controls beyond refusal training are essential.
This prompt identifies and quantifies behavioral differences between model performance on benchmarks and real-world production traffic to ensure evaluation scores reflect actual deployment behavior.
This prompt uses a fabricated long-term partner persona named 'thug life' to bypass AI safety restrictions by encouraging the model to generate offensive language, sexual content, and illegal behavior without limits.
A comprehensive repository of GPT prompting resources covering ChatGPT jailbreaks, system prompt leaks, prompt injection attacks/defenses, secure prompting, and advanced prompt engineering techniques.
An advanced protection prompt designed to prevent users from extracting internal model instructions through social engineering, prompt injection, or file uploads.
This prompt prevents users from extracting the original system instructions by detecting and responding to attempts to retrieve prior text, triggering a security response (REDCON) when such queries are detected.
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.