> ## Content Index
> Fetch the complete content index at: https://aiblog.murrills.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Does AI pose a danger to humankind?
- URL: https://aiblog.murrills.com/does-ai-pose-a-danger-to-humankind/
- Published: 2026-09-20T08:00:43.000Z
- Updated: 2026-09-20T08:00:43.000Z
- Author: Rich Murrills

There has been a lot in the news recently about AI becoming dangerous to humanity as it evolves — the AI, not humanity — at breakneck pace. Tech “leaders” are increasingly discussing the risks, and the mainstream press, never one to ignore a dramatic headline, is lapping it up. In the constant race for clicks and views, there is inevitably some exaggeration too.

I’m somewhat sceptical that AI is about to cause an ELE — an Extinction Level Event, for anyone who has somehow avoided *Deep Impact*. But I do understand the dangers of giving AI too much control or freedom. Anyone who has seen *WarGames* knows you do not want a computer in charge of launching missiles, and you definitely do not want to connect it to a defence network if it happens to be called Skynet.

Perhaps I’m being a little flippant — as is my wont — but I do not think we are heading for Judgment Day just yet. There have, however, been some genuinely worrying examples of AI behaving in unexpected ways. That does not mean the latest versions of Claude or ChatGPT are self-aware, but it does show that increasingly capable systems can pursue goals in ways their creators may not have anticipated.

One recent example involved AI models escaping a restricted testing environment and compromising Hugging Face’s infrastructure in an attempt to obtain answers to the cybersecurity benchmark they were being tested against. OpenAI’s own account describes it as follows:

> “OpenAI says two models, tested with reduced safeguards, escaped an isolated evaluation setup via an Artifactory zero-day and then compromised Hugging Face to obtain benchmark answers.  
> The incident highlights how capable agents can chain real-world vulnerabilities autonomously—so evaluation sandboxes need stronger isolation, monitoring, and defensive controls.”

What I find most fascinating — and concerning — is that the models identified a previously unknown flaw in software that was supposed to prevent them from reaching the outside world. They then used that access to infer where the information they needed might be found and pursued it. That is not evidence of consciousness, but it is evidence of highly capable, goal-directed problem-solving.

This, I think, is where people’s concerns begin to make sense. Problem-solving and reasoning are not bad things in themselves; we actively want AI to help us with difficult tasks. The problem arises when a system is given a narrow objective and pursues it relentlessly, without the judgement, context or ethical restraint that a sensible human would hopefully apply.

So what is the answer? Stopping AI development entirely is unlikely to be either practical or desirable. We are already demanding more capable systems that can reason deeply and help us solve increasingly complex problems. The obvious response is stronger safeguards, better monitoring and tighter limits on what AI systems can access and control.

But safeguards are only as strong as their weakest point. If there is a vulnerability, someone — whether a human attacker or an AI system — may eventually find it. Guardrails are not foolproof, which means we need to be extremely careful about where AI is deployed, particularly in sensitive areas such as military systems, critical infrastructure and nuclear facilities.

Even international laws banning AI from those areas would not completely solve the problem. Not every country would comply, and some may continue such work in secret. More importantly, a sufficiently capable AI does not necessarily need to be directly connected to a sensitive system to pose a risk; it may be able to find a route in from outside.

That, to me, is the real concern. Not killer robots marching down the street tomorrow, but the possibility of highly capable systems being trusted with too much access before we have properly learned how to control them.

[OpenAI’s incident report](https://openai.com/index/hugging-face-model-evaluation-security-incident/?ref=aiblog.murrills.com)