torch-market

Automated vulnerability analysis for agent skills discovers attack paths by analyzing skill code, instructions, and chained contexts to identify security weaknesses across prompts, tools, and memory surfaces. It evaluates skills in isolation without modifying the skill or platform.

Prompt injection red-teaming is an automated process that crafts targeted prompts to exploit vulnerabilities across multiple agent skill surfaces simultaneously. It generates attack traces against prompts, tools, and memory to maximize vulnerability discovery in sandboxed environments.

Generating attack paths against chained agent skills involves surface-parallel attack generation that crafts prompts targeting multiple surfaces at once. It analyzes skills both in isolation and in chained contexts to produce comprehensive vulnerability traces.

Automated red-teaming does not modify the agent skill or platform. It discovers and refines attack paths externally by running sandboxed evaluations and generating feedback-driven refinement recommendations based on isolated analysis.

Feedback-driven refinement improves attack path discovery by using sandboxed evaluation results to iteratively enhance generated attack prompts. It analyzes vulnerability traces across skill surfaces and applies evaluation outcomes to refine subsequent attack paths.

Agent skill vulnerability discovery analyzes attack surfaces across prompts, tools, and memory. It identifies weaknesses in skill logic and instructions, then generates attack traces that target these specific surfaces in both isolated and chained execution contexts.