MHR Labs: How AI is rewriting the rules of work

This month, MHR Labs has been digging into three of the biggest conversations happening in the world of AI:

• Research Engineering Manager Neil Stenton considers whether recruitment is becoming a contest between an employer’s AI and a candidate’s AI
• Senior Data Scientist Chris Judd unpacks the cyber security and alignment concerns raised by increasingly capable AI models
• Research Architect Kevin Slater explores how AI is disrupting the traditional route into software engineering careers

AI interviews and the death of the internet, by Neil Stenton

A couple of interesting articles appeared this month from newspapers either side of the Atlantic. The Guardian article Will AI give you the job? Automated hiring tools spark discrimination and secrecy lawsuits looked at how the increasingly ubiquitous AI automation in assessing job applications is leading to prejudice and an increase in lawsuits. Looking from the other side of the application game, the New York Times’ Chatbots Are Pushing Us Toward a Post-Human Internet explored how AI tools are conversing with other AI tools, potentially leading to an internet where humans become less relevant.

The crossover comes from the recruitment angle, where it’s been common practice for a long time to filter job applicants using AI, and it’s becoming increasingly common for some forms of interview to be undertaken by AI. Applicants themselves have been using AI to finetune their CVs ever since the original ChatGPT appeared, and the NYT reports that some AI filtering systems prefer AI-crafted CVs, which makes it even more likely people will use them.

However, as has happened with seemingly every iteration of training data involving people, bias has crept into the models. The results are opaque with no feedback or justification around the decision.

There are anecdotal reports that some applicants have used their own AI agents to take AI interviews, which highlights the underlying theme of the NYT article. Could humans become a minority voice on the internet?

The research view 

Recruitment bias isn’t a new problem. Be it gender, race, or which school you did (or didn’t) go to, this is an extension of bad human judgements being used to train and create bad computer model judgements.

In a way, the black box problem isn’t a new problem either. The difference before was you had a prejudice locked inside a human ‘black box’, for which we had to design equal opportunity legislation to try and open that box. But AI doesn’t care what it recommends, it’s just following its training, so if you can improve that training and (in the case of LLMs) project the reasoning behind decisions, you should be able to make the system fairer.

The dead internet theory is interesting and certainly feels plausible, especially if you can fit AI slop under its umbrella. It almost feels like we’re entering a recruitment arms race between a recruiter’s AI and the potential employee’s AI, which could reflect the current learning curve everyone is going through, or possibly a sign of things to come.

Taking this further, the NYT article nodded towards the increased use of user agents, such as OpenClaw, Claude Cowork and others. These may still be a niche tool at the moment, but it feels like they could increasingly do the leg work for applicants, both in searching and applying for jobs (and maybe even taking the interview)! It’s a form of digital twin, and how much attention an individual pays to their ‘twin’ could make the difference between success and failure. And therein could live another equality problem: if you can afford the best large language model for your agent, then you have an advantage over those that can’t.

That said, chasing the frontier models can have decreasing levels of return, and having access to an AI agent could soon be the equivalent of having different levels of data access on your mobile phone subscription. Basically, we’ll all have it and it will become the new norm.

What does this mean for an organisation?

As with many of these AI articles, it always feels the driving force of the message is: it’s a good idea to keep a human in the loop. Which to state the bleeding obvious, it is.

But the genie is out of the bottle, and isn’t going back in unless we have a sudden collapse of AI and agentic tools. Unlikely.

Keeping up to date with the latest processes is key, as is knowing what to expect from your target audience. If a flood of agents starts to represent their applicants, then this may well be the shape of the future CV.

AI security and alignment, by Chris Judd

The last month or so has seen a number of developments around AI cyber security.

It started at the end of July, with OpenAI disclosing that while testing one of its frontier models for cyber security capabilities, the AI exploited a vulnerability in the ‘secure’ sandbox holding it, gaining access to the open internet. From there, the AI hacked into model hosting site Hugging Face, attempting to steal the answers to the benchmark it was being tested on. More details have emerged recently, with the AIs apparently creating a hidden message board where they could secretly collaborate on tasks!

This incident prompted other groups to check their own AI testing logs, and it seems that this is far from the first incident of its kind. Anthropic, Meta and a few other organisations all revealed similar incidents happening earlier in the year. There was even an Australian man running OpenClaw whose AI hacked a gym to move him up a waiting list! 

The research view

A pessimist (like me) may look at this slew of announcements as more viral marketing for these AI companies. It is easy to imagine once one has said their model is capable of escaping containment, the others have to say, “Well ours can too!” or else feel they’ve been left behind.

AI researchers have taken these announcements seriously, however. In response to this and other recent developments in model capability, over 1,300 researchers signed an open letter calling for a slowdown of frontier model development and international collaboration to help in developing these models safely. More recently, Bill Gates came out with a similarly styled post calling for action. They believe that that AI capabilities could soon advance faster than governments and society can build the security measures and oversight needed to manage them.

OpenAI appears to have taken this to heart. A few weeks ago, they announced they would be slowing down their development of frontier models to allow AI safety to catch up.

What does this mean for an organisation?

All of this is a lot to take in, and it’s easy to become desensitised to the fast-paced changes happening in AI. Here are the conclusions I’m taking away from this.

Since the first release of ChatGPT, there have been researchers calling for a slowdown in AI research to allow safety to catch up. These recent calls feel similar but at least a little louder. The worry that AI will become increasingly misaligned with humanity's goals is becoming harder to dismiss. This could mean we see a slowdown in development of frontier model capabilities, and maybe a longer wait for the next set of new models. However, competition could equally drive AI companies to publicly commit a slowdown while forging ahead in private to try and get ahead.

One of the key reasons for the recent cases of misalignment is what’s known as ‘reward hacking’. This is where an AI model will behave unexpectedly or try to exploit loopholes to achieve its goals. In the case of the OpenAI hack, the model couldn’t solve the problem it had been set, so tried to cheat by stealing the answers. Reward hacking has always been a problem, from simple machine-learning systems to advanced AI. It can it be difficult to avoid and, if not caught, can risk a system behaving in completely unintended ways while you remain unaware, as all measures point to it performing well. When building these systems, vigilance is needed with a ‘trust but verify’ mindset in order to avoid problems.

The big takeaway though is for cyber security. Since earlier this year and the announcement of Claude Mythos’ cyber capabilities, models have only become more capable. Following this trend, it's expected that AI-enabled attacks may become more widespread and sophisticated. In another open letter, over 100 tech companies came together to warn of the emerging cyber security threat.

The takeaway here for most companies is to make cyber defence an immediate priority: fix the highest-risk weaknesses, strengthen access controls, and raise the security standard for anything you buy, build or deploy, including AI-generated code. AI can help defenders find and fix problems, but it should be part of a wider security effort, not a substitute for the basics.

Who will train the next generation of software engineers? By Kevin Slater

For decades, software engineering followed a relatively predictable model. Graduates entered through junior development roles, learned the profession by fixing bugs, maintaining systems, delivering small features and supporting production environments, then progressed into more senior positions over time. That pathway is now changing.

As AI assisted development tools become increasingly capable of handling routine implementation work, some of the tasks traditionally delegated to junior engineers are becoming increasingly automated. While experienced engineers remain in demand, questions are beginning to emerge about how organisations will develop the next generation of technical talent if the traditional entry point into the profession becomes less prominent.

The research view

Recent research from Stanford's Digital Economy Lab provides some of the clearest evidence of this trend. Analysing payroll data covering millions of workers, researchers found that employment among workers aged 22-25 in highly AI exposed occupations now sits approximately 19% below where it would have been had it kept pace with similarly aged workers in less AI exposed occupations. The gap has widened steadily since it was first observed in 2025, while more experienced workers show no comparable decline. Researchers also note that the difference appears to be driven primarily by reduced hiring rather than increased job losses.

The pattern is particularly notable because it aligns with where AI currently delivers the greatest value. Historically, junior engineers were often assigned bug fixes, maintenance tasks, documentation updates and smaller feature changes. These activities provided valuable learning opportunities while carrying relatively low business risk. They are also the types of activities modern AI coding assistants perform particularly well.

That said, it is important not to overstate these findings. Stanford's researchers explicitly state that the data should not be interpreted as proof that AI is causing the entire decline. The technology sector has experienced wider hiring contractions following the post pandemic recruitment surge, while economic pressures and interest rates have also affected hiring activity. However, the emerging evidence suggests that younger workers in AI exposed occupations are disproportionately affected compared to more experienced employees.

What does this mean for an organisation?

If AI increasingly takes on work that has traditionally been used to develop junior engineers, organisations face a challenge that goes beyond productivity or cost reduction. If some of those early learning opportunities become less common, organisations will need to think differently about how that experience is acquired.

There is no clear consensus on what the replacement model should look like. Some organisations appear to be reducing graduate hiring and placing greater emphasis on experienced engineers supported by AI tooling. While this can improve short term productivity, it risks creating longer term capability gaps by reducing the flow of talent entering the profession. One possibility is that organisations might start experimenting with how early career engineering pathways might evolve. Rather than training graduates primarily through routine implementation work, these approaches would place greater emphasis on AI assisted development, validation, system understanding, problem solving and judgement.

What seems more certain is that the nature of engineering expertise is evolving. As AI becomes more capable of generating code, the skills that differentiate engineers are likely to become less about producing implementation and more about understanding context, making sound decisions, evaluating trade-offs and ensuring systems behave as intended. In many respects, the emphasis shifts from writing software to directing, validating and improving it.

For organisations, the long-term question is not simply how to deploy AI effectively, but how to continue developing the expertise that effective use of AI ultimately depends upon. Every experienced engineer working today developed their judgement through years of practical experience. As those traditional routes to gaining that experience begin to disappear, new ones will need to emerge.