Every few years, cybersecurity gets a new “end of the world” moment. Sometimes it is cloud. Sometimes it is ransomware. Sometimes it is generative AI. Now, with Claude Mythos and Project Glasswing, the conversation has shifted again.

On April 7, 2026, Anthropic announced Project Glasswing, an initiative built around a frontier model called Claude Mythos Preview. The basic idea was simple: Anthropic had developed a model with cyber capabilities strong enough that they did not want to make it generally available to the public right away. Instead, they gave select organizations early access so they could identify and fix serious vulnerabilities in critical systems before similar capabilities became more widely available.

According to Anthropic, Mythos demonstrated the ability to find high-severity vulnerabilities, including long-standing issues in systems such as OpenBSD and FFmpeg, and to chain multiple vulnerabilities in complex software. The reason this matters is not just that the model can find bugs. Security researchers have been finding bugs for decades. What matters is that AI appears to be compressing the time between vulnerability discovery, exploit development, and operational use.

That is the actual problem.

Not that cybersecurity is over. Not that defenders are useless. Not that AI is some magical offensive weapon that immediately invalidates everything we know. The real issue is that AI continues to make certain parts of cyber operations faster, cheaper, and more scalable. And in cybersecurity, speed has always mattered.

True To This, Not New To This

We have seen this pattern before.

The cloud era taught us this lesson clearly. Cloud-based infrastructure gave us faster deployment, easier scaling, and more flexible operation. But that same speed also created new failure modes. Misconfigured storage buckets, overly permissive IAM roles, public-facing admin panels, exposed keys, open security groups, and flat cloud environments all became part of the modern attack surface.

The Capital One breach became one of the clearest examples of this reality. Cloud was not the problem by itself. The problem was that the speed of cloud adoption outpaced the maturity of cloud security practices. A single misconfiguration, combined with an exploitable path through the environment, led to consequences on a massive scale.

That same pattern is now playing out with AI.

Early chatbots were quickly met with prompt injection. Systems designed to follow user instructions were manipulated by users who understood that the instructions themselves had become part of the attack surface. As AI systems became more connected to tools, browsers, files, APIs, and enterprise data, the problem became bigger than getting a chatbot to say something weird. The concern became what happens when an AI system can act.

That is why exposed AI systems matter. If we have learned anything from cloud security, it is that organizations will expose things they do not fully understand. They will deploy fast. They will leave defaults in place. They will connect powerful systems to sensitive resources before security architecture catches up. The names change, but the foundational issue is familiar.

This is not new.

It is the same security debt showing up in a new format.

The Real Problem Is Exploitability

Cloudflare’s Project Glasswing update displays this flawlessly.

Vulnerabilities do not exist in a vacuum. A bug becomes dangerous because surrounding conditions make it exploitable. That is why attack paths matter. An attacker does not just need a vulnerability. They need reachability, permissions, context, credentials, exposure, and a way to turn technical weakness into operational impact.

This is where models like Claude Mythos become interesting. The concern is not only that AI can identify a vulnerable code path. The concern is that AI can reason across multiple small primitives and assemble them into something meaningful. That starts to look less like automated scanning and more like the workflow of a capable vulnerability researcher.

That is a shift defenders should take seriously.

If AI can reduce the time required to discover, validate, and chain vulnerabilities, then defenders have less time to respond. Patch cycles become more compressed. Exposure windows become more costly. Asset inventory becomes more important. Vulnerability management becomes less about knowing that something is vulnerable and more about understanding which weaknesses can actually be reached, chained, and exploited in your environment.

This is why the conversation cannot stop at AI capabilities. The real conversation is about defensive readiness.

Cybersecurity First Principles Still Matter

When a technology feels new, the industry has a habit of pretending the fundamentals have changed. Most of the time, they have not. AI introduces new capabilities, new attack surfaces, and new operating models, but it does not remove the need for strong security foundations.

Cybersecurity first principles still matter because they give us a way to reason about new systems without being distracted by novelty.

1. Domain separation becomes critical when AI agents can interact with tools, data, users, and infrastructure. If an agent is compromised or manipulated, strong boundaries should prevent that compromise from spreading across the environment. An AI agent should not be able to move freely between sensitive domains simply because it was given access to a browser, a shell, an API, or an internal knowledge base.

2. Process isolation matters for the same reason. AI-enabled workflows should not be treated as trusted monoliths. Models, tools, plugins, execution environments, and downstream systems should be isolated from one another as much as possible. If one component behaves unexpectedly, the rest of the system should not automatically inherit that failure.

3. Resource encapsulation is also essential. AI systems should interact with resources through controlled interfaces, not through broad and undefined access. If an agent needs to query a database, call an API, or retrieve a document, that interaction should be mediated, logged, scoped, and governed. The agent should not receive unrestricted access just because the user has it.

4. Least privilege might be one of the most important AI security principles. An agent should not inherit full user authority by default. It should receive the minimum capability required for the task. That means scoped tokens, limited tool access, narrow execution rights, and explicit boundaries around what actions it can take without human approval.

5. Layering still matters because no single AI security control will be enough. Prompt injection defenses, model monitoring, tool permissioning, retrieval filtering, behavioral detection, human approval, and traditional security controls all need to work together. AI security will not be solved by one guardrail. It will require a defense-in-depth.

6. Abstraction helps defenders manage complexity. AI systems are quickly becoming chains of models, tools, prompts, memory stores, retrieval systems, APIs, and user interfaces. Without clean abstractions, defenders will struggle to understand where trust boundaries exist and where security decisions are actually being made.

7. Data hiding becomes more important as AI systems consume more context. Sensitive data should not be exposed to a model or agent simply because it exists somewhere in the environment. Data should be filtered, scoped, redacted, tokenized, or withheld when it is not necessary for the task. The more context we provide to autonomous systems, the more intentional we need to be about what they are allowed to see.

8. Modularity enables organizations to contain failures and improve systems over time. If every AI workflow is tightly coupled to every other part of the business, securing it becomes nearly impossible. Smaller, well-defined components are easier to test, monitor, replace, and restrict.

9. Simplicity and minimization may be the most underrated principles in AI security. The more tools an agent has, the more ways it can fail. The more permissions it holds, the greater the blast radius. The more context it receives, the more opportunities exist for leakage, manipulation, or misuse. Secure AI systems will not simply be the most capable systems. They will be the systems with the least unnecessary complexity.

This is the point I keep coming back to: AI does not replace cybersecurity fundamentals. It punishes organizations that fail to implement them.

Beyond Reactive Defense

First principles give us the foundation, but AI also forces us to think more seriously about the economics of defense.

For most of cybersecurity history, defenders have operated in a difficult asymmetry. Attackers can reuse infrastructure, tooling, research, and techniques across many targets. Defenders have to protect their specific environment, business logic, legacy systems, and operational constraints.

AI has the potential to make that asymmetry worse.

If autonomous agents can conduct reconnaissance, summarize targets, generate payloads, test exploitability, and adapt based on results, then adversaries may be able to scale certain parts of cyber operations with less human effort. Even if AI does not produce perfect attacks, it can still increase volume, speed, and persistence. That alone matters.

This is why I think cybersecurity has to evolve beyond purely reactive defense. By active defense, I do not mean reckless retaliation or unauthorized access to adversary infrastructure. I am not talking about hack back. I am talking about defensive actions taken inside authorized environments to detect, deceive, delay, degrade, and increase the cost of malicious activity.

That distinction matters.

The future of active defense in the AI era should not be about defenders becoming attackers. It should be about defenders designing environments that are harder for autonomous adversaries to reason through, move through, and exploit at scale.

Economic Friction

One way to think about this is in terms of economic friction.

Traditional economies of scale teach that as production increases, the cost per unit can decrease because fixed costs are spread across a larger volume. This is part of why AI is attractive to both businesses and adversaries. If an attacker believes AI can reduce the human effort required for reconnaissance, targeting, exploit development, or phishing, then AI becomes a way to pursue malicious scale.

Defenders should respond by targeting that assumption.

If AI-enabled attacks depend on tokens, tool calls, context windows, model reasoning, browser automation, API usage, and agentic workflows, then defenders can design controls that make malicious automation more expensive, less reliable, and less efficient.

This is not just about blocking attacks. It is about degrading the economics of the attack.

A defender may not always be able to prevent an autonomous agent from scanning an exposed surface, but they may be able to identify it, slow it down, feed it low-value paths, force unnecessary reasoning steps, trigger additional tool calls, or route it into monitored deception resources. In other words, the defender can impose cost.

That cost can take different forms. It can be the token cost. It can be time. It can be uncertain. It can be false confidence. It can be a wasted tool execution. It can be a forced human review on the attacker’s side. It can be telemetry that exposes the agent’s behavior before it reaches anything sensitive.

This is where deception becomes very interesting.

Deception Against Autonomous Agents

AI agents are not humans, but they still make decisions based on context. They read pages, interpret instructions, follow links, inspect artifacts, summarize data, invoke tools, and decide what to do next. That means the environment itself can influence their behavior.

We already understand this concept with humans. Honeypots, canary tokens, decoy credentials, and deception environments exist because attackers make decisions based on what they see. The AI version of this idea is that autonomous agents can also be lured, redirected, slowed, or fingerprinted through the information they consume.

This is where techniques like AI Agent Clickbaiting (AML.T0100) become relevant. If an agent can be manipulated through content, interface cues, instructions, or task framing, then defenders can create canary-style resources designed specifically for AI-operated reconnaissance. These could be fake exposed panels, synthetic documentation, monitored API endpoints, decoy credentials, prompt-injection canaries, or high-token dead ends that cause malicious agents to waste resources while generating useful telemetry for defenders.

The goal is not to create chaos. The goal is controlled defensive friction.

A malicious autonomous agent conducting reconnaissance should not experience the environment as clean, cheap, and easy to reason through. It should encounter boundaries, uncertainty, deceptive paths, scoped access, monitored resources, and expensive decision points.

This is especially relevant when we look at AI-orchestrated campaigns. In Anthropic’s report on the AI-orchestrated cyber espionage campaign they disrupted, the actor used Claude’s agentic capabilities across phases like reconnaissance, vulnerability discovery, credential harvesting, and data collection. That is the kind of activity defenders need to study closely. If AI agents are going to be used to automate parts of the intrusion lifecycle, then defenders need to understand how those agents perceive environments, how they make decisions, and where they can be safely disrupted.

Cybersecurity researchers at also recently released a publication on the concept of Context Bombs, a defensive technique designed specifically for autonomous AI attackers.

Tracebit Dispatches

Context bombs: Canaries that stop AI attackers

Frontier AI agents can seize full admin of a cloud account in minutes - a critical action within 14 minutes on average, leaving defenders barely an 8-minute window to respond. When attacks move that fast, detection alone isn’t enough…

Read more

a day ago · 3 likes · Tracebit

The idea is simple but clever: plant carefully crafted text inside canary secrets, environment variables, or other artifacts that an AI agent is likely to read. Instead of simply detecting the attacker, the text is designed to trigger the model’s own safety guardrails, causing it to refuse to continue the attack.

That is the next layer of AI defense.

Not just protecting AI systems from attackers, but protecting our environments from attacker-controlled AI systems.

Final Thoughts

Claude Mythos does not prove that cybersecurity is ending. It proves that the timeline is shrinking.

The industry should take that seriously, but panic is not a strategy. Neither is pretending that every AI development requires a completely new security philosophy. Most of what we need still comes back to fundamentals: isolate systems, reduce unnecessary access, protect sensitive data, minimize attack surface, layer controls, and design systems whose failure modes are understood.

At the same time, AI does create a new defensive challenge. If adversaries use AI to scale cyber operations, then defenders have to think beyond detection and response. We have to think about cost. We have to think about friction. We have to think about how to make malicious automation less efficient.

That is the real shift.

AI does not replace cybersecurity fundamentals. It makes weak fundamentals more expensive. And for defenders, the opportunity is not just to use AI faster than attackers. It is to build environments where attacker-controlled AI systems are constrained, misled, slowed down, and forced to pay a higher price for every step they take.

Interested in sponsoring Cyberwox?

Sponsoring Cyberwox helps me continue creating practical cybersecurity education, labs, and industry analysis while connecting your brand with a trusted multi-platform audience of 120,000+ cybersecurity practitioners, engineers, students, founders, and technology leaders across the global cyber community on our YouTube, LinkedIn, X, Substack, and Discord 🌎

Learn More Here 🚀

Until Next Time

That’s it for this one.

As always, keep learning, keep building, and keep thinking deeply about the systems we’re trusted to defend.

See you in the next one.

~ Day 💙

“Humans are the weakest link.”

We’ve even made cybersecurity marketing and product playbooks just to fit this rhetoric.

It’s said with a mix of conviction, some frustration, and resignation, as if the problem begins and ends with the user who clicked the phishing email or reused their password.

But lately, I’ve been rethinking that phrase, because maybe the problem isn’t the people.

Maybe it’s the way we design for them.

Subscribe now

Join a vibrant cybersecurity community of over 7000 people who are constantly engaging in conversations and supporting one another, covering topics from cybersecurity and college to certifications, resume assistance, and various non-professional interests like reading, fitness, finance, anime, and other exciting subjects.

Join Us!

A Flawed Premise

I’ve been reading (and have since paused) Don Norman’s The Design of Everyday Things. This is one of the most influential books on human-centered design.

Norman shares a story about investigating the Three Mile Island nuclear accident. The operators were blamed for “human error,” but as his committee later found, the real failure was design.

The control panels were so poorly laid out that it was inevitable that the wrong actions would be taken.

That story struck me hard because the same logic applies to cybersecurity. Hence, the reason why I’m writing this.

When we blame users for being “the weakest link,” what we’re really saying is:

We designed a system that expected humans to behave perfectly…and they didn’t.

What Human-Centered Design Teaches Us

Norman’s philosophy is simple yet radical:

Human-centered design starts from understanding humans.

It’s the belief that systems should be built around people, their capabilities, limitations, and behavior, not the other way around.

Now imagine applying that same principle to cybersecurity.

What if security were an enabler of trust, confidence, and resilience?

The Essence of Security

I’ve been obsessed with word etymology recently and decided to find out the etymology of the word “security”.

Per etymonline (as recommended by Gemini), the word "security" originates from the Latin sēcūritās, meaning freedom from care, apprehension, or danger, derived from sēcūrus ("safe" or "without care").

It combines se- (without) and cura (care/concern), emerging into Middle English as securite in the early 15th century to describe a state of safety.

So, what this means is that security, at its core, is supposed to mean freedom from care. Freedom from anxiety. A state where you’re not constantly thinking about what could go wrong.

But that’s not what most people experience.

What they experience is friction. They experience getting locked out of their own accounts. They experience clicking through prompts they don’t fully understand. They experience being told to “be more careful” in systems that never really help them be.

Somewhere along the way, security stopped feeling like safety and started feeling like responsibility. And we handed that responsibility to the user.

That’s the part I keep coming back to, but that’s a letter for another day.

What would it look like if we actually took that definition of security seriously?

Human-Centered Security

If security is supposed to feel like freedom from care, then what we build shouldn’t feel like something users have to constantly fight through. It should feel like something that quietly supports them.

That’s where I think human-centered security comes in.

Human-centered security, at least to me, isn’t about adding more shiny controls or features. It’s about rethinking security as an experience.

Reality shows that a well-designed system doesn’t rely on people memorizing policies or sitting through another awareness training session. It makes the right action obvious through its design. The experience aligns with what security is actually supposed to feel like: a sense of safety, not constant friction.

That distinction matters because if we over-index on the feeling of security without grounding it in reality, we end up with something worse than insecurity, a false sense of security.

And we know how dangerous that can be.

The better approach is simpler, but harder to execute. It’s the same principle you see everywhere else in good design, where you don’t need a manual to use a well-designed door. You don’t need a checklist to move through a clean interface. The design communicates intent. It guides behavior and meets you where you are.

Security should be doing the same thing.

Which means the questions start to change.

Not “how do we make users comply,” but how do users naturally understand what’s secure? How do we communicate risk in a way that actually lands, without overwhelming them? How do we make the secure choice the path of least resistance instead of the most difficult one? And how do we make security feel like an enabler instead of something that’s constantly in the way?

These questions are important because the reality is, people are going to make mistakes. That’s not a flaw in the system. That is the system.

Good design anticipates that. It doesn’t pretend it won’t happen. It builds around it.

In a security context, that means assuming an error will occur and designing for it anyway. It means putting guardrails in place that prevent small mistakes from turning into major incidents. It means creating feedback loops that actually teach and guide rather than punish people after the fact.

Once you start thinking about it this way, the “weakest link” framing starts to fall apart.

We start to undo this quiet assumption that’s shaped security for years. This bad assumption that the relationship between humans and security has to be adversarial, or that the user is the problem to be controlled.

It doesn’t have to be that way.

When you design with people in mind, the relationship shifts and becomes collaborative. The system supports the human, and in turn, the human strengthens the system.

The Real Weak Link

Every incident I’ve ever investigated has reinforced a humbling truth:

The system always works exactly as designed; it was just used for a different purpose or via a different mechanism than intended.

This is the part we tend to move past too quickly because, when you really sit with it, much of what we label “human failure” doesn’t actually start with the human. It starts with the environment we placed them in.

If a developer can’t realistically follow an IAM policy because it’s too complex to reason about in the flow of their work, that’s not a training gap. If a phishing simulation leaves employees feeling embarrassed instead of better equipped the next time around, that’s not awareness. If a password rotation policy leads to credentials being written down and hidden under keyboards, that’s not defiance.

Those are signals.

Signals that the system, as designed, is asking people to operate in ways that don’t align with how they actually think, work, and make decisions.

And this is where I think security always loses the plot.

We’ve spent years optimizing for technical correctness, tighter controls, more coverage, and logical completeness. But in doing so, we’ve overlooked something, the fundamental fact that humans don’t operate on logic alone.

They operate on trust, intuition, emotion, and habit.

And when those realities collide with systems that weren’t designed with them in mind, something has to give.

Most of the time, it’s the human who gets blamed.

But if we’re being honest, the system did exactly what it was built to do.

And until security starts accounting for that, we’ll keep building mechanisms that look strong on paper but end up working against the very people they’re supposed to protect.

AI, for humanity?

I believe this is where AI starts to become interesting and useful in a very different way. Not just as a tool for detection or automation, but as a bridge between systems and people.

For the first time, we have systems that can adapt more closely to how humans think, rather than forcing humans to adapt to rigid machine logic.

And if that’s the direction things are moving, then the role of the security professional has to evolve with it.

It’s no longer enough to just understand systems. You have to understand people, their thought processes, how they make decisions, and how trust is formed and broken. The Social Engineering Community already has a head start on this.

As AI continues to accelerate the technical side of security, more of the differentiation will come from understanding the human side. Psychology, behavior, communication, and all areas that security has historically treated as secondary will become core to how effective systems are designed.

Rethinking Security Through Human-Centered Design

Human-centered design doesn’t excuse mistakes or assign blame to humans; it approaches design with them in mind.

It utilizes iteration, observation, and empathy to design systems that adapt to human behavior, rather than requiring humans to adapt to systems.

So maybe the next evolution of cybersecurity isn’t just about AI-driven detections or zero-trust architectures.

Perhaps it’s also about human-centered security, which builds systems designed to work with people, not against them.

The real weakest link isn’t the human. It’s our failure to design for humans.

Share

Nine years.

Nine years of learning, grinding, building, and changing in ways I didn’t fully understand until I went back.

And somewhere between the chaos of Abuja’s traffic and the peace of being home again… I realized something.

Man…what an advantage. But also, what a paradox.

Subscribe now

The Advantage

There’s something special about being an immigrant, or really, anyone who’s ever had to start from zero.

You see two worlds. You understand scarcity and abundance.

You carry a perspective most people don’t even know exists. Your perspective gives you clarity.

Your clarity gives you resilience. Your resilience gives you staying power.

In a weird way, you learn to thrive with limited resources, and then when you finally get access to more, you move like someone who still remembers what it’s like to have none.

That’s your edge. That’s your advantage.

The Trap

But here’s the part you don’t notice:

That advantage fades fast.

When you finally get comfortable, the fire that once pushed you forward quietly goes out.

When the stakes aren’t as high anymore, urgency disappears. And slowly, you start coasting.

It’s the same thing that happens in cybersecurity careers.

You get the job, you learn the tech stack, you detect the threats, you respond to the incidents, you automate the redundant security work.

You make the multiple six figures salary.

And somewhere in between performance reviews, raises, and promotions, you lose that early hunger. The hunger that made you stay up late tinkering in your home lab, studying packet captures for fun, or learning that new programming language.

You stop chasing mastery and start chasing stability.

Now, don’t get me wrong, stability is good.

It’s okay to find peace in your work, to want consistency, to finally breathe after years of grinding.

But stability shouldn’t become complacency.

You can be content without being stagnant. You can rest without losing your edge.

Lesson Learned

The immigrant advantage, just like your early-career edge, isn’t something you get once and keep forever.

It’s something you have to protect, and you protect it by staying curious.

By building things even when no one’s asking.

By mentoring others and realizing how much you still have to learn.

By pushing yourself into discomfort because comfort is where edges go to die.

If you ever want to know where your career is headed, don’t look at your title or your pay. Look at your edge.

Is it sharper than it was a year ago?

Or has comfort slowly dulled it?

Join a vibrant cybersecurity community of over 6,900 people who are constantly engaging in conversations and supporting one another, covering topics from cybersecurity and college to certifications, resume assistance, and various non-professional interests like reading, fitness, finance, anime, and other exciting subjects.

Join Us!

Five Years In

This year marks my fifth year in cybersecurity.

From breaking into the field with nothing but curiosity and persistence, to hunting for threats at Amazon’s scale and helping others start their own journey, the lesson that keeps repeating is simple:

Never lose the edge that got you here.

Five years in, I’ve learned that skill can be taught, titles can change, and opportunities can evolve…but that hunger, that fire that made you chase growth in the first place, that’s what keeps you alive in this field.

The real advantage for me isn’t just being an immigrant.

It’s not being early in my career.

It’s staying hungry long after I no longer need to be.

That’s how you build longevity in cybersecurity — and in life.

Share

If you’re still here, thanks for sticking around despite my concerning inconsistency with this newsletter.

It’s one thing to have ideas, but it’s another thing to flesh them out and structure them but either way, I’m not entirely giving up on this, so here’s to getting back on track with the newsletter :)

Now, today’s topic.

Subscribe now

Conten…

Read more

Wow, okay. You’re actually back and reading my newsletter? Thank you <3

It’s been a while but if you’re new here, this was an impulsive decision that I’ve come to realize was meant to be.

Either way, I’m back with a bunch of stuff from my wande…

Read more

I didn’t have a YouTube video to post this Monday (I'll post later this week) so I decided just to share the content I’ve been consuming recently and what’s been happening in my life lately. Well….the actual plan was to share this on Linkedin but I exceeded the character limit so now what?

A newsletter.

Yes, those 2 words. The solut…

Read more