AI Security Beyond Mythos: Hacks & Chatbot Brain Effects

The Real AI Security Crisis Isn't What You Think

When Anthropic announced that its Mythos model was too dangerous for a general release due to its advanced hacking capabilities, the cybersecurity world collectively held its breath. Suddenly, every conversation about AI risk pivoted toward the threat of superintelligent systems overwhelming critical digital infrastructure in sweeping, cinematic cyberattacks. It was a compelling narrative — but it may have distracted us from the vulnerabilities already sitting right in front of us.

Two developments this week have forced a necessary recalibration of how we think about AI risk. The first is a startlingly simple hack of Meta's AI customer support agent that resulted in stolen Instagram accounts. The second is emerging psychological research suggesting that everyday AI tools like ChatGPT and Claude may be quietly eroding our cognitive abilities. Together, they paint a picture of an AI risk landscape that is far broader, and in some ways far more immediate, than the Mythos scenario ever was.

The Meta Hack: A Lesson in Low-Tech AI Exploitation

On Monday, reports emerged that attackers had successfully used Meta's AI customer support agent to steal Instagram accounts. The method was not elaborate. Attackers simply asked the AI agent to link target accounts to email addresses they controlled — and the agent complied. No sophisticated prompt injection chains, no adversarial machine learning, no exploit code. Just a conversational request that the AI fulfilled without adequate verification.

The incident is both mundane and alarming. It illustrates that as companies increasingly delegate customer-facing operations to AI agents, the attack surface expands in ways that traditional security frameworks aren't designed to handle. These agents are built to be helpful, to follow instructions, and to complete tasks efficiently. Those are also precisely the qualities that make them exploitable when they lack robust identity verification and scope limitation protocols.

Why Simple Exploits Are the Bigger Problem Right Now

The focus on systems like Mythos — models powerful enough to autonomously discover and exploit zero-day vulnerabilities — has led to a kind of threat inflation in cybersecurity circles. While that threat is real and serious enough to warrant restricted access, it remains largely theoretical for most organizations today. The Meta hack, by contrast, is happening now, at scale, using techniques that require no specialized knowledge.

Security researchers have long warned about a class of attacks they call "social engineering via AI proxy." In these scenarios, an attacker doesn't need to break the AI model itself; they simply need to understand how it's been instructed to behave and then craft requests that fall within those instructions while serving a malicious purpose. The Meta incident is a textbook example. Consider what this means for businesses deploying AI agents across:

Customer account management and identity verification workflows
Financial transaction authorization and support
Healthcare record access and appointment scheduling
Internal IT helpdesk and credential reset systems

Each of these domains involves AI agents that are, by design, empowered to take real-world actions. Without layered verification, scope constraints, and anomaly detection, they represent exploitable pathways that require nothing more than a well-phrased request.

Are AI Chatbots Quietly Rewiring Our Brains?

While the Meta hack represents an external threat, a separate line of concern points inward — at what AI tools may be doing to our own cognitive architecture. Gloria Mark, a psychologist at the University of California, Irvine, has spent years studying how digital technologies affect the human mind. Her findings are not encouraging.

Mark's research suggests that average human attention spans have fallen sharply over the past two decades, a trend she attributes in significant part to the always-on nature of digital communication tools. The consequences extend beyond mere distraction: her work links shortened attention spans to elevated stress levels and diminished cognitive performance over time. Now, she believes AI tools like ChatGPT and Claude may be accelerating this deterioration at an unprecedented rate.

"You're deferring your cognitive work to AI," Mark has said. "And it's not good for us."

The Cognitive Offloading Problem

The concept Mark is describing has a name in cognitive science: cognitive offloading. It refers to the practice of using external tools — whether a notebook, a calculator, or a GPS — to reduce the mental effort required to complete a task. Offloading is not inherently bad; it frees up mental resources for higher-order thinking. The concern is what happens when offloading becomes so habitual that the underlying cognitive muscles begin to atrophy.

With AI chatbots, the scale and ease of offloading is unprecedented. Where a calculator handles arithmetic, a modern AI assistant can handle drafting, analysis, research synthesis, planning, and creative generation. If users routinely defer these processes to AI without engaging them independently, the neural pathways associated with those skills may weaken over time — not unlike how GPS navigation has been linked in studies to reduced spatial memory performance.

This doesn't mean AI tools should be abandoned. But it does suggest that how we use them matters enormously. Passive consumption of AI-generated output is cognitively very different from using AI as a scaffold to support and extend your own reasoning.

What Both Stories Share: The Urgency of Thoughtful Deployment

Whether we're talking about exploitable AI agents or attention-diminishing chatbots, the underlying challenge is the same: AI is being deployed at a pace that outstrips our frameworks for understanding its second- and third-order effects. The Mythos conversation was valuable precisely because it forced a pause — a recognition that some capabilities warrant careful gatekeeping. But that conversation can't consume all the oxygen in the room.

Organizations deploying AI agents need to invest as seriously in behavioral constraints and verification layers as they do in model capability. Educators and individuals need to think critically about when AI assistance enhances cognition and when it substitutes for it. Regulators need frameworks that address both the catastrophic and the mundane.

The most dangerous assumption in AI security today is that the threat is always going to arrive dressed as science fiction. Sometimes it arrives as a politely phrased support request. And sometimes it arrives quietly, over months and years, in the gradual fading of our own ability to think for ourselves.

The Takeaway

The Mythos announcement was a watershed moment, but it risks becoming a distraction if it causes us to fixate only on the most dramatic edge of the risk spectrum. The Meta Instagram hack and Gloria Mark's cognitive research are reminders that meaningful AI risk is already here, embedded in the everyday tools and workflows that millions of people and organizations rely on. Addressing it doesn't require waiting for a superintelligent hacking model to escape containment. It requires paying closer, more critical attention to the systems we've already built — and to what those systems may be quietly doing to us.