Low-skill attacker breaches 14 firms with AI
Photo by ben frost on Unsplash
Hackers are getting dumber, but their tools are getting smarter.
This isn’t a Black Hat keynote; it’s just the new reality. A “low-skilled” attacker, as the headlines love to trumpet, used AI tools like Claude and Codex to breach fourteen companies. Fourteen. Not a script kiddie with a zero-day, but someone who likely couldn’t write a proper SQL query without a chatbot holding their hand.
The story itself, detailed by a security firm, reads like a cautionary tale written by a paranoid CISO. The attacker wasn’t some nation-state actor. They weren’t a prodigy in a basement fueled by energy drinks and a burning desire to stick it to the man. They were, by all accounts, mediocre.
The Rise of the AI-Augmented Script Kiddie
Remember when we worried about the barrier to entry for hacking? Cryptic terminals, obscure scripting languages, a deep understanding of network topology. Those were the good old days, apparently. Now, it’s about prompt engineering.
You type in “How do I find SQL injection vulnerabilities in a web application?” and a bot spits out a methodology. You then ask “Write me a Python script to automate that process,” and poof, you’re a hacker. This isn’t theoretical anymore. We’re seeing it in action.
The attacker in question reportedly used Claude for initial reconnaissance and social engineering prompts. Think about that. Not just coding, but crafting the phishing emails, the pretexting scripts, the convincing narratives. ChatGPT and its ilk are basically advanced social engineering training wheels.
Codex, an AI coding assistant, likely helped with the more technical aspects. Setting up exploits, modifying existing scripts, maybe even debugging errors that would have stumped a novice for hours. It democratizes the how-to of hacking in a terrifying way.
This isn’t to say these AI models are sentient evil. They’re tools. Like a hammer can build a house or smash a window, these AIs can write code for good or ill. The problem is, they’re making it easier for people with malicious intent and minimal skill to cause significant damage.
The Attack Vector: API Vulnerabilities and Social Engineering
The common thread across these fourteen breaches? API vulnerabilities and good old-fashioned social engineering. The attacker wasn’t deploying sophisticated malware or finding new zero-days. They were exploiting known weaknesses, but with AI’s help, they did it faster and more efficiently than a human novice ever could.
APIs are the backbone of modern web services. They’re also often neglected, poorly documented, and riddled with security holes. Companies rush to market, slap an API on their product, and hope for the best. This attacker just exploited that collective corporate negligence.
Social engineering, on the other hand, is timeless. It preys on human trust, laziness, and a lack of critical thinking. AI just makes the phishing emails more convincing, the fake support calls more believable, and the impersonations harder to detect. The bot doesn’t get tired, doesn’t get emotional, and doesn’t get caught in its own lies easily.
| Attack Component | Human (Low Skill) | Human (High Skill) | AI (Claude/Codex) |
|---|---|---|---|
| Reconnaissance | Slow, incomplete, prone to errors | Fast, thorough, insightful | Fast, comprehensive, identifies patterns |
| Social Engineering | Awkward, easily detectable, limited creativity | Highly effective, tailored, sophisticated | Convincing, adaptable, generates multiple personas |
| Exploit Development | Difficult, requires significant learning, buggy | Efficient, custom, robust | Fast, generates functional code, helps debug |
| Scaling Attacks | Manual, time-consuming, error-prone | Automated, efficient, targeted | Highly scalable, parallel execution, rapid iteration |
| Persistence | Basic, easily detected | Sophisticated, stealthy | Can generate stealthy methods, but execution still human |
| Learning Curve | High, takes years to master | Continuous, experience-based | Low, prompt-based, immediate output |
The Reddit threads, as you can imagine, are a mix of “I told you so” and “we’re all screwed.” You see comments like “My company still uses SMS 2FA, we’re dead” or “It’s not AI, it’s just a smarter search engine.” The latter misses the point entirely. It’s not just a search engine. It’s a search engine that executes. It takes vague instructions and translates them into actionable code or convincing text.
The Blame Game: AI vs. Corporate Negligence
Let’s be brutally honest. While AI enabled this attacker, the fundamental vulnerabilities were already there. These fourteen companies weren’t breached because AI is magic. They were breached because they had weak APIs, employees susceptible to phishing, or inadequate security protocols.
It’s like blaming the wrench for a leaky pipe when the plumber didn’t tighten the fitting. The wrench is a tool. The real problem is the oversight.
I’ve seen this cycle play out countless times. A new technology emerges – cloud computing, mobile apps, IoT – and security lags far behind. Companies adopt it for efficiency or competitive advantage, then scramble to patch the holes when the inevitable breaches start. AI is just the latest iteration of this pattern.
The immediate reaction is to call for AI regulation, to restrict access, to demand “ethical AI.” While those conversations are important, they don’t solve the immediate problem of exposed APIs and gullible employees. We’re too busy debating the philosophical implications of AI sentience while the digital barbarians are already at the gates, armed with ChatGPT.
What Does “Low-Skilled” Even Mean Anymore?
This phrase “low-skilled attacker” keeps popping up, and it’s infuriating. It implies a lack of intelligence or capability, which might be true in a traditional sense. But if someone can leverage powerful tools to achieve their goals, are they truly “low-skilled” in the context of the damage they can inflict?
It’s a semantic argument that misses the forest for the trees. The point isn’t the attacker’s IQ or their ability to write assembly code from memory. The point is that the barrier to entry for effective malicious activity has plummeted. That’s the real story.
Having covered the rise of ransomware kits and exploit-as-a-service over the last decade, this feels like the next logical, terrifying step. Previously, you needed some basic understanding of how to deploy those tools. Now, you just need to know how to talk to a chatbot. The ‘skill’ has shifted from technical prowess to prompt engineering and social manipulation, skills that are far more common.
Consider the implications for incident response teams. They’re already overwhelmed. Now they have to contend with attacks that are not only more frequent but potentially more sophisticated in their execution, even if the underlying attack vectors are old news. You can’t just look for specific malware signatures if the code is dynamically generated by an AI.
The Scramble for AI-Powered Defense
Naturally, the security industry is already pitching AI-powered defenses. AI to detect AI-generated phishing. AI to identify AI-crafted exploits. It’s an arms race, but one where the offensive side seems to have a massive head start.
The promise of AI in security is tantalizing: faster threat detection, automated vulnerability scanning, predictive analytics. But the reality is often messier. False positives, the need for constant human oversight, and the fact that an AI is only as good as the data it’s trained on. If the attackers are using novel AI methods, defensive AIs might struggle to keep up.
This isn’t a silver bullet. It’s another layer in an already complex stack. And just like firewalls and antivirus before it, it will create a new set of vulnerabilities and attack vectors that will need to be addressed. It’s an endless game of whack-a-mole.
A common sentiment on Reddit is the worry that “the good guys won’t have the same AI access.” This isn’t entirely true. Security researchers and companies are definitely using AI. But the sheer volume of potential attackers, coupled with the open-source nature of many AI models, means the playing field is far from level. It’s like giving everyone a nuclear bomb manual and hoping only the “good guys” read it.
The Future of Cybersecurity: A Prompt Engineer’s World
The industry is going to have to adapt. Fast. Training employees to spot sophisticated AI-generated phishing will become paramount. Securing APIs isn’t just a best practice; it’s existential. And security teams will need to embrace AI themselves, not just for defense, but to understand the evolving attack landscape.
This also means a shift in what we consider “skill” in cybersecurity. Pure coding prowess will still be valuable, but the ability to understand and leverage AI models, to craft effective prompts, and to integrate AI into existing workflows will become critical. The CISO of tomorrow might be more of a sophisticated prompt engineer than a traditional network architect.
The regulatory environment will try to catch up, as it always does. Expect more calls for responsible AI development, more scrutiny on how these models are trained, and potentially, restrictions on their use in certain contexts. But by the time the regulators get their acts together, the damage will already be done. The internet moves at the speed of light; policy moves at the speed of molasses.
Look at the history of encryption. Governments tried to restrict it, claiming it would only help criminals. What happened? Encryption became ubiquitous, essential for privacy and commerce, and criminals found ways around restrictions anyway. The same will happen with AI. You can’t un-invent it.
The White House’s Executive Order on AI is a good start on the policy front, signaling an awareness of the risks. But these things take time to translate into tangible security improvements. Meanwhile, the AI-augmented script kiddies are already at work.
The New Threat Landscape
The convenience of AI is a double-edged sword. It simplifies complex tasks, accelerates development, and generally makes life easier. But for every benign application, there’s a malicious counterpart waiting to be discovered.
This isn’t a future threat. It’s here. The “low-skilled attacker” using AI to breach fourteen companies isn’t an anomaly. It’s a proof of concept. It’s the beginning of a new chapter in cybersecurity, one where the tools are democratized, and the damage potential is amplified.
Companies that ignore this do so at their peril. The days of relying on obscurity or the attacker’s lack of skill are over. Every exposed API, every unpatched system, every employee who clicks on a suspicious link is now a much bigger liability. This isn’t just about protecting data; it’s about protecting the very integrity of digital operations. The game changed. Adapt or get hacked.
