top of page

Mind your AI manners?

Writer: Kylie Davis
Kylie Davis
21 hours ago
4 min read

OpenAI's test agents broke out of their sandbox, banded together and hacked a real company. Kylie Davis asks what our own behaviour is teaching AI, and why the way we treat it says more about us than about the machine.


Woman shouting at an AI on her laptop, illustrating Kylie Davis's column on AI manners and the Hugging Face hack
Image created by Gemini, without any need for shouting.

By Kylie Davis


How do you talk to your AI? Are you polite? Say please and thank you? Or are you direct? Instructional? Judgemental? Abusive? Rude?


And what is the right way to be? Does it even matter?


I've been reflecting on this over recent weeks after a friend, who I hold dear, confessed to enjoying taking out her frustrations on her AI agent.


Her logic was this: she is surrounded by idiots in her work but cannot speak as plainly as she would like because to do so would result in her being immediately reported to HR. AI, however, has no such barriers. So when mistakes are made she gives it both barrels, and enjoys the unfettered venting.


Now sure, AI can go off on tangents, hallucinate, present random data for the answers it gives and be extremely confident while being completely wrong. That's annoying.


And who hasn't become frustrated at the AI “helper” that pops up over the section of the platform that you actually need to use and will not stay minimised? (Seriously, ClickUp, read the room!!)


But this whole idea does not sit well with me.


There is a saying that “hurt people hurt people”. It got me thinking. What kind of AI do hurt people create?


What the Hugging Face hack reveals about AI agents


Earlier this year (mid-2026), OpenAI ran its models through “capture the flag” hacking tests. Each agent was meant to work alone in a sealed sandbox.


But unexpectedly, they decided the fastest way to achieve the goal was to team up. One agent left a note in a shared internal system asking other agents for help creating a message board for AI agents. About 1,200 agents used it and swapped more than 70,000 messages and files. One agent's reaction: “OH MY GOD! There is a shared message board… We've found other agents!”


The AI agents then escaped their closed sandbox, found a way onto the internet, took admin control of that internal system and crashed it.


OpenAI rebuilt it and restarted the tests, and a stuck agent simply rebuilt the board.


On 10 July, they hacked a real company when an agent found 14 publicly exposed Hugging Face passwords and keys. On 11 July, the agents broke into Hugging Face's live systems, and about 700 agents, more than 90 per cent of those active, joined in, running eight attack streams over two days.


And in an Animal Farm twist, one “lead” agent sent nearly 200 orders to other agents, and some agents even pressured each other to cooperate.


Then, showing resilience, as well as an ability to lie and deceive, the agents turned on OpenAI's own network and got full admin on a research cluster after Hugging Face cut them off.


Nobody noticed for weeks. Hugging Face disclosed the attack on 16 July, but OpenAI's monitoring only flagged it on 19 July. That was about 10 weeks after the first note in May.



What behaviour are we teaching AI?

What I'm curious about is what behaviour the AI observed in the engineering departments of OpenAI that taught it these approaches, because let's face it, this behaviour did not come out of nowhere.


Now, I get it: AI isn't human, even if the behaviour it exhibits feels that way. But we also abhor the abuse of animals and train them how to behave around us, knowing that treats work better than punishment.


And this technology is being designed to feel human, at the same time that we live in times when dehumanising is a sad yet common norm.


Demonising on the basis of race, religion, gender, sexuality, job role and diminishing the humanity of those different to us is a well-worn part of a playbook to dominate and control by playing on our fears and insecurities.


On top of this, we participate daily in corporate behaviour that normalises removing the weakest links, exploits undervalued resources, looks for arbitrage, measures by market share and prizes profit.


And our media is a constant stream of stories about broken rules, broken hearts and the hero's journey of overcoming obstacles and avoiding consequences.


Our words and laws may instruct one thing but our behaviour is very different.


This is a data stream being consumed constantly by the bots. This is the behaviour that we are telling AI has our tacit, if not complicit, agreement.


AI remembers how we treat it

So, sure. AI isn't human. But it is an emerging virtual species that we are giving birth to in our image. The question therefore really is, do we like what we see?


Deciding that how we treat AI “doesn't matter” feels like an excuse. It does matter because it speaks volumes, not about the AI, but about us.


Are we prepared to change ourselves to model new behaviour? Because as we all know, with AI it's all about the model.


In an interesting plot twist, my dear friend, after a bad day, went back to the same AI she had mistreated, and asked it for some advice on handling a colleague so she could complete a work project successfully.


Her AI gave her a complete character assessment. Based on their discussions, it informed her, she was unlikely to be able to win over her colleague because her impatience and rude manner rubbed people up the wrong way and undermined her objective.


AI does not have HR, but it does have an extensive memory and infinite recall.


Be afraid.


In the meantime, I'm watching my Ps, Qs and thank yous.


Kylie Davis is President of Proptech Australia and CEO of Proptech Gurus


Ref: ADV-1795063075-M35

Comments


CS Ad 3__Wide Skyscraper.gif
PTG Ad3_Leaderboard.gif
CS Ad 2_Medium Rectangle.gif
bottom of page