An LLM that will help you build a nuclear weapon

AI safety guardrails in LLMs like GPT and Claude are criticized as liability-driven theater rather than genuine safety measures. The author, attempting to penetration-test their own sandbox, found mainstream models unhelpful for legitimate security research. They discovered 'abliterated' models—open-weight models with refusal mechanisms surgically removed—which proved far more useful for tasks like enumerating privileged access tokens. The piece argues that corporate censorship of AI outputs, analogous to 3D printer legislation, harms legitimate users while barely inconveniencing bad actors, and that OpenAI's 'Trusted Access for Cyber' program is too restrictive for independent developers. The author contends that knowledge is multipurpose and that corporate liability concerns, not genuine safety, drive these restrictions.