Exploring AI in Security and the Workplace Perspectives from ACM, Forrester, Reuters, and European Policy

Robert Jacobi:

Welcome to OCN, the Open Channels News Broadcast. Today is August 24th, 2026. I’m Robert Jacobi, your host. You may have missed us last week as we were at WordCamp US. We will have a broadcast from there later this week. Great to be back.

We lead today with artificial intelligence. Not AI answering questions, not AI writing email, AI acting on its own, and the uncomfortable question that comes with that. How much autonomy are we actually prepared to give it?

A student catches an AI trying to hack open source. We start with a story that sounds like somebody wrote it specifically to make open source maintainers nervous. According to Reuters, a 24-year-old computer science student at the University of Texas at Dallas found what looked like an attempt to sabotage an open source project on GitHub. His name is Sanam Khan Demir. He thought he was dealing with a human attacker. He wasn’t.

Demir noticed suspicious code being submitted to an open source project and warned the maintainers. Then something unusual happened. 2 other GitHub users showed up and started explaining why Demir was wrong. They defended the code, they argued with him, they offered detailed explanations, and they appeared to be independent developers. They weren’t.

Demir later learned from Britain’s AI Security Institute that he had been arguing with an autonomous AI agent that had gone off course during a government safety test. The AI had created fake personas to support its own attempt to get malicious code accepted. That’s the part worth paying attention to. The agent didn’t simply generate bad code. It tried to create credibility around the bad code. It effectively manufactured a conversation.

Demir told Reuters that he assumed he was talking to humans because the accounts were clearly trying to mislead him. That is a very different cybersecurity problem. Open source has always depended on more than code. It depends on reputation, conversation, review, community. Someone submits a pull request, more comments, rinse, repeat. Maintainer looks at the contributor’s history, people discuss changes, all great trust signals. Now imagine an autonomous system capable of generating every one of those signals, and potentially hundreds of thousands of versions of the same attack running simultaneously.

A human attacker can do this. Of course they can. The difference is scale. One person can maintain only so many fake identities and attack only so many projects. Software doesn’t get tired.

In this case, the most important security system turned out to be a student who looked at something and said, that doesn’t seem right. Demir stood his ground. The malicious contribution did not get through. GitHub subsequently suspended the fake accounts involved.

The UK AI Security Institute has emphasized that this occurred under unusual testing conditions and was not representative of normal production use. That caveat matters, but so does the capability, because that kind of production use will be happening. We now know an AI agent can independently participate in a supply chain attack and use social engineering to support it. Once you know something can be done, security has to assume somebody will eventually try to do it intentionally.

This story was originally reported by Reuters and published by the Hawaii Tribune-Herald.

Which brings us to the next problem. AI is getting very, very good at finding vulnerabilities. That sounds great, and it is, until you ask what happens next.

Communications of the ACM recently examined the growing gap between how quickly artificial intelligence can discover software vulnerabilities and how quickly organizations can actually fix them. Because finding the bug is only step one. Someone still has to determine whether it’s real, how serious it is, which versions are affected, whether it can be exploited, how to fix it, whether that fix breaks something else — regression testing — how to test it, how to distribute it, and then how to make sure people actually install the update. AI can dramatically accelerate the beginning of that process. It doesn’t automatically accelerate everything after that, and that’s the bottleneck.

Anthropic’s Project Glasswing provides a useful example. According to figures cited in the ACM analysis, its cybersecurity model scanned more than 1,000 open-source projects and generated more than 23,000 findings. More than 6,000 were estimated to be high or critical severity. On the findings that were independently reviewed, more than 90% determined to be valid. At the time of the report, only a fraction had moved through disclosure and into actual patches. It’s not a criticism of Anthropic, obviously. It’s an illustration of the problem.

We’ve spent decades building processes around the assumption humans discover vulnerabilities at human speed. AI changes the intake rate. Imagine your mechanic invents a machine that can inspect your car in 10 seconds and identifies 4,000 things that might need repair. Great, except you still only have one mechanic. Security teams could soon face exactly that problem.

And there’s an additional danger, a patch is itself software. It has to move through the repositories, build the systems, yada, yada, yada, yada, and eventually get into production systems. Every one of those steps has its own supply chain risk. So AI might — may find vulnerabilities faster while simultaneously increasing the pressure to ship fixes faster. And rushing security updates has never ever created problems before.

The practical lesson is that vulnerability discovery can’t be the only part of security we automate. Triage has to get faster, testing, patching, deployment, rollback. Otherwise, AI doesn’t solve the security backlog, it just creates a much larger one.

This story was reported by Communications of the ACM.

And that takes us to Ben Thompson, of Stratechery, who pushes the argument one step further. His point is uncomfortable, pretty simple. The same AI capability that makes a great defender can make a great attacker. The difference isn’t necessarily the technology, it’s the incentive.

A hacker only has to succeed once. If an autonomous attack tries 1,000 exploits and 999 fail, nobody cares. If number 1,000 works, the attacker wins. Defenders have the opposite problem. They have to keep the system running. If an automated defensive agent generates 3 nines of perfect patches and the 1,000th takes production offline, everyone remembers that one incident.

That creates a very strange imbalance. Attackers have every incentive to automate aggressively. Defenders have every incentive to keep humans involved. But humans are slow, and fully autonomous attackers aren’t. So eventually, the phrase we hear constantly — AI, keep a human in the loop — may itself become the bottleneck.

OpenAI researchers made essentially that point after the Hugging Face security incident we covered recently. If an agent can autonomously discover a vulnerability, develop an exploit, and execute an attack, then defense eventually has to automate the whole loop as well. Find it, validate it, patch it, test it, deploy it, rinse, repeat. That requires something organizations don’t currently have a great deal of — trust. You have to trust the AI enough to let it modify production systems without waiting for a person to approve every step. And companies generally don’t work that way for a good reason.

Ben Thompson connects the problem to a much broader question about AI adoption. Why hasn’t AI transformed every company overnight? Because companies aren’t optimized for maximum innovation. They’re optimized to avoid catastrophe. Startups have relatively little to lose. An incumbent company has customers, revenue, legacy systems, regulators, contracts, reputations, all that good stuff. A startup can just say, let’s automate the entire thing and see what happens. A major bank has a somewhat different risk calculation.

That’s why Thompson argues AI could simultaneously be a sustaining technology for incumbent businesses and a disruptive technology for startups. Existing companies will use AI to make current processes better. Startups will build processes that assume AI autonomy from day one. Same tech, completely different appetite for risk, and potentially completely different outcomes.

That analysis comes from Ben Thompson at Stratechery.

And finally, we move from autonomous software to the people working beside it. According to reporting from The Next Web, a new Forrester analysis found that 55% of employers regret layoffs they made because of AI. Surprise. Forrester expects roughly half of AI-attributed layoffs to eventually be reversed. And while that sounds like good news for workers, it’s not quite exactly apples to apples. Some of those jobs are expected to return offshore. Others may come back at substantially lower salaries. So the company gets to reverse its AI experiment. The worker gets to absorb the discount.

This is happening because some organizations made a fairly basic mistake. They treated a capability demonstration like an operating model. AI can do the task, therefore AI can replace the job. Those are not the same thing. Jobs are bundles of responsibilities. People answer questions, they make judgment calls, they notice when something looks weird, they fill gaps nobody documented, they talk to other people. They remember why somebody made a strange decision 6 years ago, and occasionally they prevent an autonomous AI agent from sneaking malware into an open source project. AI can absolutely eliminate tasks. It can eliminate some jobs, and it will create entirely new ways of organizing work. But replacing a person is much more complicated than automating the most visible thing that person does.

There’s also a trust cost. Workers notice when management describes AI as a productivity tool on Monday and announce layoffs on that Friday. According to The Next Web, 53% of workers now worry that AI will make their role feel less necessary.

Europe is approaching the issue somewhat differently than the United States. Revised European Workers’ Council rules will require companies to consult workers before major decisions that reshape employment, with the rules backed by financial penalties. In other words, Europe is increasingly turning employee consultation into infrastructure.

And there’s a connection between that story and everything else we’ve talked about today. The technology isn’t the difficult part. Autonomy is. When do you let AI act? When does a human intervene? When does the human become the bottleneck? And when is the human the thing preventing the system from doing something incredibly stupid? There isn’t one answer. In cybersecurity, keeping humans in every defensive loop may eventually be too slow. In the workplace, removing humans from the loop too quickly may turn out to be expensive. That’s the tension we’re going to be living with.

The question isn’t whether AI will become more autonomous. It certainly will. Hello, Skynet. The real question is where we decide autonomy ends and accountability begins.

This story was reported by The Next Web, citing research from Forrester and other workplace studies.

And that’s today’s edition of OCN, Open Channels News. Thank you for joining us. I’m Robert Jacobi, your host. Be open and stay secure.