Identify AI/ML security weaknesses using adversarial examples and model extraction attacks.
npx skills add https://github.com/0X6C7879/aegissec --skill ctf-ai-ml
Please help me install this Agent Skill. Skill: ctf-ai-ml Source: https://github.com/0X6C7879/aegissec/tree/main/skills/ctf-ai-ml Command: npx skills add https://github.com/0X6C7879/aegissec --skill ctf-ai-ml
CTF participants need practical techniques to assess and exploit AI/ML systems in controlled scenarios.
Run a practice AI/ML attack against a target model using the provided examples to observe defense gaps in a safe, controlled environment.
High-intent search queries and answers about installing and using this skill.
AI/ML attack techniques for CTF challenges include adversarial examples, data poisoning, model extraction, membership inference, encoder collisions, LoRA exploitation, and LLM attacks. These methods map model vulnerabilities to practical exploits for retrieving flags.
To reproduce adversarial machine learning attacks, use guided payloads and references to model attack papers provided by the skill. These resources help you execute staged exploits and observe defense gaps safely within controlled test environments.
Yes, LLM attacks are a core scenario. The skill provides practical exercises and actionable payloads specifically designed to assess LLM security weaknesses, understand risks, and demonstrate vulnerabilities in controlled CTF scenarios.
No specific dependencies are required. The skill focuses on providing actionable guidance, workflows, and example payloads mapped to reference papers, allowing you to apply model extraction and data poisoning techniques using your preferred environment.
The best way to map an ML target's vulnerabilities is applying a broad set of scenarios like adversarial examples and membership inference. This identifies practical weaknesses and helps design staged exploits to demonstrate system risks.
Adversarial examples should not be used outside controlled, safe test environments. They are designed for CTF scenarios to observe defense gaps and understand risks; deploying them against production models without authorization is unsafe.
pluginagentmarketplaceAnalyze training data for poisoning and backdoor attempts in ML workflows.
jeffpeckyIdentify security issues and generate reports with mitigation strategies for codebases.
wuyoscarBenchmark NVIDIA Aegis AI Content Safety classifier across six harm categories.
rainoftimeGenerate fuzzing tools that create random inputs to discover software bugs.
