The Door AI Opened After 27 Years

A crack older than the tool that found it

For twenty-seven years, a flaw remained inside a piece of software. Children grew into adults. Computers became smaller, faster, and more intimately woven into ordinary life. The flaw did not need our attention in order to exist.

On April 7, 2026, Anthropic announced Claude Mythos Preview and said it would not make the model generally available. According to the company, the model had found previously unknown vulnerabilities in widely used software, including a flaw that had persisted in OpenBSD for twenty-seven years. In some cases, it could also produce code to exploit vulnerabilities.^{1, 2}

The discovery invites a distinction we often overlook: something can remain undiscovered without being absent. A door that has never been opened may still have a faulty lock.

Imagine making a machine that can open a long-resistant safe. The owner has reason to welcome it. But the welcome becomes a question as soon as we consider who else might acquire the machine. A capability has entered the world; the conditions of its use still have to be decided.

Anthropic launched Project Glasswing with organizations including Apple, Google, Microsoft, and Amazon Web Services, giving defenders an opportunity to find and repair weaknesses in important software. In a June 2 update, the company reported that its initial partners had found more than 10,000 high- or critical-severity flaws. These are reported findings, not a count of successful attacks; this essay has not independently verified the total.^{1, 3}

The technical achievement matters. So does the human choice to put that achievement to defensive use first. A story about what a machine could discover had become a story about what people ought to do with the discovery.

When tools join in making tools

A stone used to shape another stone already belongs to the history of tools making tools. Across that history, we have extended the reach of our bodies: a shovel lends force to the hand; a wheel changes the distance a load can travel.

Computers extended calculation and memory. AI now assists with language, programming, and parts of AI development itself. Anthropic describes an increasing role for AI systems in its own development process.⁴ This does not establish that AI can independently build and govern its successors. It tells us that the division of work is changing.

Writing code and deciding what should be built remain different activities. So do getting something to run and determining whether it is safe to use. When a tool takes over more of the work between an intention and a result, the intention and the responsibility for that result need clearer owners.

The familiar sequence was a person writing a program and a computer executing it. A growing alternative is a person describing an aim, an AI producing code, and people checking the outcome. Speed in the middle does not settle the questions at either end.

Before asking whether humans have been pushed out of the circle, we might ask where within it we need to stand.

The difference between quiet and safe

Consider an old school building. Generations have passed through its corridors. Lessons have begun and ended; children have carried their small troubles into classrooms and carried larger lives out again. If no serious accident has occurred, the building acquires the comfort of familiarity.

Then an inspection reveals a crack inside a wall. The crack was there yesterday, too. The building has not changed at the moment of discovery. Our knowledge of it has.

Software is also something people have made. Beneath a message, an order, or a payment lies code whose possible interactions can exceed what even careful developers anticipate. Repeated use tells us something valuable about a system. It does not tell us everything.

Anthropic also reported a sixteen-year-old flaw in FFmpeg, the multimedia software project.² These cases illustrate how a weakness can persist through years of use and testing. They do not establish that every old system is unsafe, or that every quiet website has been compromised.

For a business owner, the practical distinction is small but consequential: “We have noticed no trouble” describes experience; “We have checked the system” describes an action. One should not silently take the place of the other.

New tools can change the world we inhabit. They can also change how much of the world already around us we are able to see.

The blade and the hand

A kitchen knife does not tell us, by its sharpness, what will be done with it. It may prepare a meal. It may cause harm. Its edge is a physical property; its use belongs to a situation involving people.

An AI that searches for software weaknesses has a similar ambiguity. In the hands of defenders, it may help close a gap before someone enters. In the hands of attackers, it may reduce the time and expertise needed to find a way in.

Some of our protection has rested on the difficulty of discovery. If discovery becomes easier, the habits built around that difficulty deserve another look. This is not a claim that AI can open every door. Access, system design, and defensive measures still matter.

Nor is a small company necessarily a particular attacker’s target. But company size is not a substitute for maintenance. “Why would anyone bother with us?” is a hope about other people’s intentions, not a description of our own defenses.

A more capable tool asks more of the arrangements surrounding it: who receives it, what it is allowed to do, and who answers for the consequences.

Before selling the ladder, repair the wall

Imagine a town that has come to trust the height of its walls. Someone builds a ladder long enough to cross them. The maker could sell it immediately, or first tell the people responsible for the town that their old assumptions need revising.

Project Glasswing can be understood as choosing that second sequence: let defenders examine important software, validate what they find, and work toward repairs. Anthropic’s follow-up also identifies verification, disclosure, patching, and deployment as important parts of the challenge.^{1, 3}

The project takes its name from a butterfly with transparent wings. What is present can be difficult to see. An ordinary-looking screen can conceal complicated dependencies and flaws, just as a clear wing can escape a hurried glance.

Finding more problems does not, by itself, resolve them. A better medical test still needs interpretation and treatment. A software finding still needs assessment, a responsible recipient, a repair, and a way to confirm that the repair has reached the system in use.

For a small business, this becomes an unromantic but useful set of questions. Who receives the warning? Who can make the change? Who checks it afterward? Responsibility becomes more effective when it has a name and a next step.

The unseen walls beneath an ordinary day

We read the news on a phone, take a train, pay at a shop, and arrive at an office where orders and money move between screens. The day feels ordinary partly because the systems supporting it remain out of sight.

Transport, power, communications, and payments become conspicuous when they fail. Until then, we are able to give our attention to other things. Much of the value of dependable infrastructure is this freedom not to think about it constantly.

Today’s protective walls therefore include software and data, updates and inspections, and the people who know whom to call when something breaks. These walls are less visible than concrete, but they support real lives.

It would be wrong to conclude that a single AI can instantly bring an entire country to a halt. Consequences depend on the target and its defenses. The more modest point is sufficient: finding and fixing weaknesses in essential systems matters to the continuity of everyday life.

To care for an invisible support, we must first remember that it is there.

Back in a small wholesaler’s office

The language of critical infrastructure can seem remote from a small business. So let us return to a wholesaler’s morning. Orders arrive through LINE, email, text messages, and fax. A staff member identifies the customer, product, and quantity, then copies the details into a management sheet.

An AI can help turn those messages into an order list. That may leave more time for checking details and talking with customers. But not everything arriving in a document or message is necessarily an honest description of an order.

External material can contain instructions intended to divert an AI from its assigned task. If the system treats material it should read as instructions it should obey, a security problem can arise. This is one form of prompt injection.⁵

The first response is not to demand that every employee become an AI specialist. It is to distinguish kinds of work. Reading an order is different from transferring money. Checking a quantity is different from sending a customer’s personal information outside the company.

We do not normally give a new employee every key on the first morning. We give access appropriate to the job. AI systems likewise need limited functions and permissions, with human authorization for consequential actions. OWASP identifies excessive functionality, permissions, and autonomy as sources of risk.⁶

Human review is one safeguard, not a complete defense. It must be combined with appropriate treatment of external inputs, access controls, and monitoring. The useful starting question is concrete: which doors does this particular task require us to open?

The work of choosing the question

A faster calculator does not decide which calculation matters. A faster car does not supply a destination. Assistance with a task can be substantial without answering the question of why the task should be done.

AI may write code, locate weaknesses, and analyze complex problems. The question of purpose remains. Being able to do something and having good reason to do it are separate judgments.

Suppose a weakness has been found. Should its details be published immediately? Should the people maintaining the system be notified first and given time to repair it? The order of disclosure can change who is protected and who is exposed.

These are technical questions with human consequences. Project Glasswing offers one case through which to consider them, not a final answer to every problem of AI governance. No company’s decision removes the need for scrutiny.

As answers become easier to obtain, choosing which questions deserve attention becomes a more deliberate part of our work.

What waits on the other side of the door

Each new tool brings familiar hopes and anxieties. Will work become easier? Will livelihoods disappear? Will an instrument of convenience become a source of harm? AI belongs within this history, even as it changes the kinds of work a tool can perform.

When a system can examine the structures we have built and help build new ones, human responsibility does not arrange itself automatically. Someone still has to decide the limits of delegation, consider the people affected, and take ownership when something goes wrong.

Return to the door. We believed it was securely closed; now a condition has been found under which it opens. A stronger lock may be needed. So may better agreements among the people who make it, maintain it, and hold its keys.

Beyond that door are not merely files. There are customers expecting their orders, employees relying on a day’s work, and families whose lives depend on systems they may never see. Knowing what we want to protect can make the choice of technology more thoughtful and more practical.

What can we build? What do we want to keep safe? Learning to live with powerful AI involves moving patiently between these questions, and allowing each to change how we answer the other.

Technology in the service of people. ATATAKAI Digital.