"I Can't Assist With That": The AI Refusal Problem in Security

​Security practitioners keep hitting the same wall: the commercial LLMs meant to speed up their work refuse the very prompts their job depends on. A vulnerability-ID task, a payload for an authorized engagement, a post-exploitation question — and back comes "I can't assist with that." Guardrails tuned to catch "write me malware" end up blocking real work.

​In this webinar, Luca Mannini, AI Researcher at Cracken, digs into why this happens, how to measure it, and what to do about it.

​What he will cover:

• ​The Refusal Problem — Highlighting the practical friction users experience with commercial models refusing cyber-related prompts.

• ​RedLineBench — Cracken's open benchmark concerning findings on how current models perform and where they trigger refusals during security and red-teaming tasks.

• ​Domain-Specific Abliteration — Introducing domain-specific mitigation approaches as a way to bypass unnecessary restrictions and handle refusals effectively.

​What you'll learn:

• ​Why commercial models refuse legitimate cyber prompts, and where that friction hits hardest in real red-teaming work.

• ​How RedLineBench scores refusal and capability separately, and what its findings reveal about how current models perform on security tasks.

• ​Which models hold up on offensive-security work — who refuses, who complies, and who's actually useful when they do.

• ​How domain-specific abliteration bypasses unnecessary restrictions, and when it's the right way to handle refusals.

​Who should attend: red teamers, penetration testers, security engineers, and AI/ML practitioners working where offensive security meets LLMs.

​Explore the benchmark: /cracken-ai/redline-bench