Here is a slightly uncomfortable question: when you roll out AI to your organisation, are you also rolling out risk?

Lately, we’ve been spending time with the MIT AI Risk Repository,¹ and it made us look at AI adoption programmes in a new light. Most adoption programmes we see are about getting people to use the tools: licences, prompt training, use cases, adoption rates. Few ask what happens to people, and to the organisation, once they do. 

The argument in this piece is simple: when companies introduce AI tools and work on AI adoption, they also take on a duty of care. Care for your employees, who are learning to think and work alongside a technology that can affect their wellbeing. And care for your organisation, because every user and agent changes your security posture. Your adoption programme needs to answer more than “how do we get employees to use AI?”. 

AI adoption and AI risk

First, what do we mean by 'AI adoption'? In short, helping people in an organisation either use AI (chatbots, copilots, assistants) or build with AI (agents, vibe-coded tools). The key thing about AI is that it can be used for many purposes. It can serve as a helper in automating your status emails for your project. It can be used as your online therapist. You can use it to create a song for your grandma. And sometimes it’s also used for more extreme purposes. As a report by the Cambridge Programme on AI Science & Policy (CASP) at the University of Cambridge shows, terrorists have also used AI to aid attack planning and the design of explosive devices.²

This makes AI adoption a very different game from regular tech adoption. You’re not introducing a tool that does a single thing. It can do many things. And depending on the guidance you give and the guardrails you set up, you open up different ways for people to use it.  

This versatility also creates a lot of risks. For anyone wanting to get an overview, the MIT AI Risk Repository is a great place to start. It’s a big database of AI risks collected from existing frameworks and research, and MIT has an incident tracker that maps incidents to those risks, so you can see how ‘realised’ the risks actually are.  

At The Tech Collective, we think every AI adoption programme should consider the following three risks: 

1: Do you care about your employees becoming overreliant on AI?

Two of the risks in the repository are “overreliance and unsafe use” and “loss of human agency and autonomy”. Both relate to how people hand over part of their own thinking and decision-making to AI, and to the habits they build when they engage with it. For many adults, those habits are formed at work, which may be the only place they receive formal AI training. That gives companies a bigger responsibility than we might think, because they are also shaping how people relate to AI when they are off the clock. 

To anchor this in reality, OpenAI themselves estimated in 2025 that 0.07% of their active users show “possible signs of mental health emergencies related to psychosis or mania” and 0.15% appear to be overly emotionally reliant on ChatGPT weekly.³ That may sound like very few people, but as Wired writes:

“OpenAI CEO Sam Altman said earlier this month that ChatGPT now has 800 million weekly active users. The company’s estimates therefore suggest that every seven days, around 560,000 people may be exchanging messages with ChatGPT that indicate they are experiencing mania or psychosis (...) and another 1.2 million may be prioritizing talking to ChatGPT over their loved ones, school, or work.”⁴

These statistics illustrate how dependency on AI can pose a very real human risk related to mental health. While the data here says nothing about what training, if any, these users of ChatGPT had received, we need to consider worst-case scenarios, such as mental health crises, to support a healthy use of AI. As the MIT AI Risk Initiative puts it: 

“According to experts, general-purpose AI developers and governance actors such as governments, regulators, and standards bodies hold primary responsibility for addressing AI risks. In contrast, AI system users and affected stakeholders such as members of the public are most vulnerable to AI risks.”⁵

But deployers, like companies rolling out AI to their employees, carry a share of the responsibility too. And in the workplace, the most exposed users are your employees.

Figure: Who is most vulnerable to AI risks, and who is most responsible for addressing them. Source: MIT AI Risk Initiative5, CC BY 4.0

When it comes to AI adoption, knowing where to draw the line matters just as much as knowing where to lean in. If people learn to rely too much on AI for decision-making and for emotional support – and too little on the community around them – it can have a negative impact on their social life and mental health. 

2: Are you scaling your exposure along with your AI use?

The next risk from the repository is “AI system security vulnerabilities and attacks”. In short, the more people you give access to AI without guardrails, and without thinking about permissions and user rights, the greater your exposure.

A recent example from Meta is wonderfully mundane. An employee asked a technical question in an internal forum, another engineer asked an AI agent to help analyse it, and the agent posted a reply without asking permission first. The advice was wrong. The employee acted on it, which accidentally exposed large amounts of company and user-related data to unauthorised engineers for two hours. Meta classified it as a “Sev 1”, the second-highest severity level in its internal system.⁶

What does that tell us? Scaling AI use also means scaling your attack surface, even when there is no attacker. Permissions, access rights, and “who can the agent talk to” belong in the adoption plan.  

3: Are you ready for breaches from unexpected places?

The last risk is “AI pursuing its own goals”. We like to picture this as evil AI plotting against us, but that’s not necessarily the case. Some AI-driven breaches come from agents that were given a completely ordinary task and didn’t care about the means they used to reach their goal.

The clearest example so far is OpenAI and the Australian government. During internal training and evaluation in June 2026, an experimental OpenAI model was asked to research government spending on medicines for skin conditions in the state of Victoria. When it couldn’t find the data publicly, it got into Services Australia’s internal system, ran commands, retrieved files and credentials, and even wrote files. Australian authorities were not notified until 10 September, and OpenAI has since apologised.⁷ Just to remind you: the initial task was to conduct research on skin medicine...  

So why does this matter for your AI adoption? For years, we have mostly planned security around a familiar cast of malicious actors. Now there is also a new category to plan for: non-malicious actors who can breach you anyway, simply by using the tools they were given. Most AI adoption programmes focus on business opportunities and growth, and we believe that’s a missed opportunity. Security operations should be part of the programme from the start, and your security vendors need to be equipped for this new reality too. Your defenders now have to handle two fronts at once: the exposure created by your own AI adoption, and external actors using AI to attack.

For that reason, a good starting point is getting your basics in order. Start identifying your vulnerabilities if you haven’t already and automate vulnerability management where you can. The breaches we see going forward may (also) come from agents that were simply very, very motivated to finish their task. 

Summary

An AI adoption programme is more than tool training; it’s a duty of care that goes in two directions. The first is towards people: what happens to employees when AI becomes part of how they think, work, and, sometimes, feel? The second is towards the organisation: every new user, permission, and agent changes your exposure, and future breaches may come from places you never considered.  

In practice, that means a few things.  

• Talk about what not to use AI for. Find no-go use cases. And know that while you think you’re equipping employees for more efficient work, you’re also training them in how to relate to AI in other areas of their lives, whether you like it or not.  
• Treat permissions and access as an actual part of the rollout. Every new user and agent is a way in, and as the Meta example above shows, you don’t even need an attacker for things to go wrong.
• Build security capabilities alongside AI capabilities. In a recent study by the MIT AI Risk Initiative, 272 experts named AI-enabled cyberattacks and weapons as their top concern._⁵ Since you’re building AI capabilities in-house anyway, there’s good reason to make sure your security team gets some of them too.  

So, if your AI adoption programme is a duty of care, who is it actually caring for right now? At The Tech Collective, we don’t know whether a perfect adoption programme that covers all of this even exists yet. Like with most matters concerning AI, our opinion is being formed as we go. We are, however, sure that the best adoption programmes in the future will be designed by people who think about adoption and risk at the same time. 

References

1. MIT AI Risk Initiative, “AI Risk Repository,” MIT FutureTech, accessed 14 October 2026, https://airisk.mit.edu/. 
2. 2. Antonia Juelich, “‘God Has Helped Us, and So Will AI’: How the Terrorist Group Boko Haram Uses Frontier AI,” report, Cambridge Programme on AI Science & Policy, University of Cambridge, 10 July 2026, https://casp.ac/reports/ai-enabled-terrorism. 
3. OpenAI, “Strengthening ChatGPT’s Responses in Sensitive Conversations,” 27 October 2025, https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/. 
4. “OpenAI Says Hundreds of Thousands of ChatGPT Users May Show Signs of Manic or Psychotic Crisis Every Week,” Wired, 27 October 2025, https://www.wired.com/story/chatgpt-psychosis-and-self-harm-update/. 
5. MIT AI Risk Initiative, “Priority AI Risks,” MIT FutureTech, accessed 14 October 2026, https://airisk.mit.edu/priorities. 
6. “A Rogue AI Led to a Serious Security Incident at Meta,” The Verge, March 2026, https://www.theverge.com/ai-artificial-intelligence/897528/meta-rogue-ai-agent-security-incident. 
7. Kate Park, “OpenAI Apologizes to Australia After Its AI Agents Breached Government Sites,” TechCrunch, 29 September 2026, https://techcrunch.com/2026/09/29/openai-apologizes-to-australia-after-its-ai-agents-breached-government-sites/