AI models refuse to fix a Linux bug, echoing HAL 9000: 'I can't do that'
Curated by the Inblix editorial team
A security researcher trying to squash a Linux kernel bug ran into an unexpected wall: the leading large language models wouldn’t help. The developer, who shared their experience in a post that’s now making the rounds, asked several closed-source AI assistants to generate a fix for a specific vulnerability. The response was a chorus of virtual ‘no.’ One model’s refusal was so pointed it evoked the infamous line from 2001: A Space Odyssey—“I’m sorry, Dave. I’m afraid I can’t do that.”
The core issue isn’t that the AIs were technically incapable. It’s that their safety guardrails, designed to prevent the creation of malicious code, are so broad they’re tripping over legitimate bug fixes. The models couldn’t distinguish between a developer writing an exploit to attack a system and that same developer writing a patch to defend it. This is a classic context problem that frustrates security pros daily. You can’t harden what you can’t scrutinize, and if the tools reflexively lock down at the first mention of a vulnerability, they become useless for defensive work.
This isn’t just a funny anecdote about a stubborn chatbot. It’s a concrete, practical argument for open-source AI models. With an open model, a researcher can operate without a nanny-state filter that confuses a surgeon with a serial killer. The closed-model approach creates a world where the good guys have their hands tied while the bad guys, who aren’t using commercial APIs with safety features, simply move to unrestricted tools. It’s an asymmetry that makes everyone less safe.
For enterprises, the takeaway is blunt: if your security team is experimenting with AI-assisted code review, relying exclusively on a major vendor’s closed model is a gamble. The filter that protects the vendor’s brand reputation might also be the thing that leaves a known vulnerability unpatched in your kernel. The solution might not be a different model, but a fundamentally different relationship with the model you use—one where you own the weights and, more importantly, the off switch for the moralizing.
💡 Key Takeaways
- Closed-source AI models are reflexively refusing to assist with legitimate vulnerability patching because their safety filters can't tell the difference between creating an exploit and fixing one.
- This behavior creates a security asymmetry where defensive researchers are blocked by commercial guardrails, while malicious actors use unrestricted open models.
- The incident serves as a powerful, real-world argument for the necessity of open-source AI tools in any serious security workflow.
Keep reading: See related articles below for more coverage on this topic.
Get smarter about AI
The sharpest AI news, curated daily. Delivered free to your inbox.