Invisible Unicode Attacks, CrowdSec Source Leak, and the AI Coding Boom Explored

Robert Jacobi:

Welcome to OCN, the Open Channels News broadcast. Today is September 21st, 2026, and I’m your host, Robert Jacobi. Today, 3 stories about what happens when software gets easier to manipulate, easier to expose, and easier to create. Microsoft finds attackers hiding invisible characters inside ordinary-looking phishing emails. CrowdSec discovers that private source code had been copied months before anyone noticed. And the New York Times asks a very good question: if AI can write software this quickly, where are all the great new applications? Different problems, same underlying issue. The cost of doing things with software is collapsing. The cost of knowing whether those things are good, safe, or trustworthy is not.

We start with something you literally cannot see. The Next Web reports that Microsoft researchers found a large phishing campaign using invisible Unicode characters to try to evade email filters. This technique is called ASCII smuggling. Unicode includes a block of tag characters that software can read but people generally cannot see. Security researchers have already demonstrated how those characters can hide prompt injection instructions from humans while still feeding them to AI systems. But this campaign used the same trick for something much more old-fashioned. Yep, spam. Spam, spam, spam.

Attackers inserted invisible characters into words such as “funding.” A person still sees the word normally, but a security system doing literal matching may see 2 different fragments instead of one recognized word. And at scale, this was not a small experiment. Microsoft observed weekday volumes between roughly 1 million and 2.37 million messages during the campaign.

The good news is that it mostly did not work. Microsoft says more than 99% of the messages were caught by other defensive layers, including sender reputation, domain checks, authentication, brand impersonation detection, classifiers, and even OCR that looks at the rendered messages the way a person would. There’s a nice security lesson in that. The attacker tried to become invisible, and the invisible characters themselves became a signal. Microsoft says that those tag characters are so unusual in normal email that their presence can now be treated as a high-confidence indicator.

That is why layered security matters. If your entire defense depends on matching one word, splitting that word can beat you. If you’re also looking at identity, reputation, behavior, URLs, domains, rendering, and context, one clever encoding trick becomes much less clever. And there is an AI lesson here too. The same hidden character technique can attack both traditional filters and AI agents. Normalize the content before the machine reads it, because if software can see something that the human cannot, you should probably know what it is seeing.

Now to CrowdSec, where the security company has disclosed that private source code was copied from its GitHub in May and surfaced publicly in September. CrowdSec says it learned about the leak on September 16th after an archive appeared on a breach forum. Its investigation traced the exposure back to May 22nd of ’26. The company says the archive included more than 130 public repos along with private repos.

Sounds bad. It is bad. But the distinction between source code exposure and infrastructure compromise matters. CrowdSec says its infrastructure and databases were not accessed, its software was not modified, and its build pipelines were not compromised. The company traced the incident to credentials associated with a former employee’s laptop. Access was revoked on May 25th, but by then the source code had apparently already been copied. CrowdSec says it rotated keys and credentials, investigated the archive, and worked with GitHub to reconstruct what happened.

This is one of those incidents that makes security people uncomfortable, because it’s not a story about somebody forgetting cybersecurity exists. CrowdSec is a cybersecurity company. Its team includes people with red team and penetration testing backgrounds. They know the rules, and they still got caught.

It’s useful to remember: security expertise does not create immunity. Companies change, teams shrink, employees leave. Access persists a little longer than it should. Credentials sit on laptops. Exceptions get made because somebody wants to finish some work. It’s a normal operational reality. The security problem is that attackers only need one of those normal moments to line up.

There’s also an open source wrinkle here. Part of CrowdSec’s code is supposed to be public, part is not. So a source code leak does not tell you very much by itself. You have to ask which code, what credentials were exposed, whether production systems were reached or pipelines altered, and whether anyone really needs to take action. CrowdSec says the exposure was limited to source code and that its infrastructure was not compromised. The larger lesson is simple: offboarding is a security control, not just paperwork.

And finally, the New York Times opinion section asks a question that gets more interesting every month. AI can code, so where are the killer apps? The argument is not that AI coding tools do not work. They obviously do. People who could never build software before can now describe an idea in plain language and get something running. Experienced developers can generate code much faster. The barrier between having an idea and having a prototype has collapsed, but that does not automatically mean the world gets better software.

Because writing code was never the only hard part. You still need a useful idea. You still need product judgment, design, to understand your users, and to decide what not to build. And once everybody can create software cheaply, simply creating software stops being much of an advantage.

That may be the most interesting part of the AI coding boom. For decades, code itself was scarce. If you wanted an application, you needed developers, time, money, infrastructure, and usually a decent reason to spend all of those things. AI changes that equation. Software becomes abundant, and when something becomes abundant, value moves somewhere else. Maybe it moves to distribution, trust, proprietary data, community, brand, or simply knowing which problem is actually worth solving.

There’s a temptation to assume that because AI can produce 10 times more code, we should expect 10 times more innovation. Those are definitely not the same thing. A word processor lets us produce more words. It did not automatically produce more Hemingway. A camera phone created billions more photographs, and it did not make everybody Annie Leibovitz. AI coding may do the same with software. We’re going to get much, much more of it. The interesting question is how much of it anybody actually wants.

So that’s the thread for today. Attackers use invisible characters because software does not always perceive information the way humans do. CrowdSec’s source code exposure shows how one access control exception can outlive the moment that created it. And AI coding is making software dramatically easier to produce without making judgment, security, or product sense dramatically easier to produce.

That’s a trade-off of cheaper creation. You get more good things, you also get a lot more bad things, more mediocre things, and things that someone still has to really look at. The bottleneck moves. It used to be, can we build it? Increasingly, the question is, should we trust it? Should we use it? And did anybody actually need it?

And that’s OCN for September 21st, 2026. I’m Robert Jacobi. Be open and stay secure.