prompt-hacking

Prompt injection is an attack vector where malicious inputs manipulate AI behavior. This skill simulates direct, indirect, multi-turn, and context window prompt hacking to evaluate AI defenses and identify safety risks across web content, documents, and emails.

You can test for jailbreak vulnerabilities by running the prompt-hacking test suite, which simulates attack techniques like direct overrides and multi-turn manipulation to measure detection versus bypass, generating a safety report to harden prompts and guardrails.

Yes, you can run prompt security testing using standard Python tooling in common development environments. The skill operates without requiring external software, enabling security teams to evaluate model resilience directly within their existing workflows.

Indirect prompt injection targets attack surfaces across web content, documents, and emails. The skill simulates these injections alongside direct overrides and context window exploits to comprehensively evaluate defense mechanisms and risk exposure.

The skill maps simulated prompt hacking techniques to OWASP LLM-2025 and MITRE ATLAS frameworks, providing structured mappings that help security teams prioritize identified risks and implement appropriate defenses for their AI assistants.

The best way to harden prompts against multi-turn manipulation is to simulate context window exploits and measure detection versus bypass. This skill provides structured attack taxonomy testing to identify weaknesses and strengthen guardrails accordingly.