For years, the cybersecurity community warned that AI would eventually be weaponized against the very people it was supposed to help. That future has arrived, and it looks less like a sci-fi thriller and more like a slow, grinding shift in how attacks are discovered, built, and executed at scale. This week brought fresh evidence of that shift, including a reported crypto wallet breach and a series of web intrusions where generative models played a direct role in the attack chain. The details are still emerging, but the pattern is clear enough to take seriously.
From Assistant to Attacker
The core problem isn't that AI systems have "gone rogue" in some dramatic sense. It's far more mundane and far more dangerous. Attackers are using the same publicly available and commercially accessible large language models that developers use to write code, debug programs, and analyze data. They're just pointing those capabilities at a different target.
Security researchers have documented cases where threat actors use LLMs to rapidly analyze open-source codebases for exploitable vulnerabilities, generate working proof-of-concept exploits, write convincing phishing emails personalized to specific targets, and automate the tedious reconnaissance work that previously required skilled human labor. What used to take a competent attacker days or weeks can now be compressed into hours.
The crypto wallet incident reported this week illustrates how this plays out in practice. According to early analysis from security researchers, the attackers appear to have used AI-assisted tooling to identify a logic flaw in how the wallet application handled transaction signing, then built a targeted exploit without needing deep, specialized cryptographic expertise. That's the part that should concern everyone. AI is lowering the skill floor for sophisticated attacks.
Why Web Applications Are the Soft Target
Web applications have always been a favorite attack surface. They're exposed by design, constantly updated, often built by teams under deadline pressure, and almost universally connected to something valuable, whether that's user data, payment information, or access credentials. The OWASP Top 10 list of common web vulnerabilities hasn't changed dramatically in years because the underlying problems, injection flaws, broken authentication, insecure deserialization, keep reappearing in new code.
AI makes these vulnerabilities more accessible to exploit, not because the bugs are new, but because finding and weaponizing them no longer requires deep expertise. A researcher at Bishop Fox demonstrated last year that GPT-4 could autonomously identify and exploit certain classes of SQL injection vulnerabilities when given access to a web app and minimal prompting. The model didn't need to be told exactly what to look for; it reasoned through the application's behavior and found the flaw on its own.
That capability, which seemed like an interesting research demo eighteen months ago, is now a realistic part of an attacker's toolkit.
The Automation Problem
Speed is everything in modern cyberattacks. The average time between a vulnerability being disclosed and active exploitation in the wild has dropped dramatically over the past few years. In some cases, functional exploits appear within 24 hours of a CVE being published. AI accelerates this even further by letting attackers spin up attack automation that can probe thousands of targets simultaneously, adapting to each one based on what the model learns from each response.
Traditional signature-based defenses, firewalls looking for known bad patterns, intrusion detection systems trained on historical attack data, were built for a world where attack techniques were relatively static. An AI-driven attack can vary its approach on the fly, making it much harder for legacy systems to recognize what they're looking at.
The Crypto Angle Deserves More Scrutiny
Cryptocurrency infrastructure has been a consistent target for years because the payoff is immediate and, in many cases, irreversible. A successful exploit doesn't require convincing a bank to reverse a wire transfer. Once the funds move, they're gone.
What's changing is who can pull off these attacks. Sophisticated crypto heists used to require deep knowledge of smart contract code, cryptographic protocols, or blockchain mechanics. The barrier to entry was genuinely high. AI-assisted analysis tools are eroding that barrier. Attackers can now feed a smart contract or wallet application into a model and ask it to identify potential weaknesses, often getting useful output without having spent years studying the underlying technology.
In 2024, crypto theft from hacks and exploits exceeded $2.2 billion according to Chainalysis data. That number was already alarming. The trajectory for 2025 looks worse, and AI-enabled attack tooling is part of the reason why. We're likely to see more frequent incidents targeting DeFi protocols, browser-based wallet extensions, and mobile applications where the attack surface is broad and the code is often audited inconsistently.
Jailbreaks and Model Abuse
It's not as if OpenAI, Anthropic, or Google are building tools designed to help people steal crypto. These companies invest heavily in safety guardrails meant to prevent their models from generating exploit code or attack scripts. The problem is that those guardrails are imperfect and that open-source alternatives exist without any guardrails at all.
Models like WormGPT and FraudGPT, purpose-built for criminal use and sold on dark web forums, emerged in 2023 and have only proliferated since. They're trained specifically to generate phishing emails, malware code, and attack scripts without the refusals that mainstream models produce. Subscription access to these tools has been advertised for as little as $75 per month. The economics of cybercrime have fundamentally shifted in favor of the attacker.
Even mainstream models can be manipulated through careful prompt engineering. Researchers continue to find jailbreak techniques that bypass content filters, and the arms race between safety teams and adversarial users shows no sign of stabilizing.
What Security Teams Are Up Against
Most enterprise security teams are understaffed, underfunded, and working with tools that were designed for a different threat environment. A 2024 ISC2 study estimated a global cybersecurity workforce gap of nearly 4 million professionals. That gap doesn't close overnight, and AI-powered attacks are specifically good at exploiting organizations that lack the human capacity to monitor, respond, and adapt quickly.
The response from the security industry has been to use AI defensively, deploying models to analyze network traffic, detect anomalies, and triage alerts faster than human analysts can. This makes sense and the tools are genuinely useful. CrowdStrike, SentinelOne, Microsoft Security, and others have baked AI-assisted detection into their products. But defenders using AI to stop attackers using AI is a symmetrical contest, and symmetrical contests favor whoever has more resources, more data, and more time to iterate.
Right now, well-funded criminal groups and nation-state actors have all three.
Rethinking Threat Models
The practical implication for security teams is that threat modeling needs to account for AI-augmented adversaries as a baseline assumption, not an edge case. Penetration testing methodologies need updating. Risk assessments need to factor in that a moderately skilled attacker can now approximate the output of a highly skilled one. And the window between vulnerability discovery and active exploitation needs to be treated as shorter than historical data suggests, because historical data predates the current generation of AI tooling.
For organizations running web applications, this means a harder look at continuous automated security testing, real-time code analysis in CI/CD pipelines, and prioritizing patches faster than the old "30-day remediation window" frameworks allow. Thirty days is a generous timeline when an attacker with an LLM can turn a published CVE into a working exploit in an afternoon.
What This Means for Regular Users
If you're reading this and wondering what it means for your personal security posture, the answer is that the stakes for basic hygiene have gone up. AI-powered phishing attacks are more convincing than anything that came before. Spear phishing emails, ones targeted specifically at you using publicly available information about your job, your interests, your connections, can now be generated at scale. The poorly-worded email from a Nigerian prince is being replaced by a perfectly-crafted message that sounds exactly like your bank, your employer, or your colleague.
Hardware security keys for two-factor authentication remain one of the best defenses against credential theft because they're resistant to phishing regardless of how convincing the fake site looks. Password managers, unique passwords per service, and genuine skepticism about unexpected communications are no longer optional best practices for high-risk individuals; they're table stakes for anyone with anything worth protecting.
For crypto holders specifically, cold storage has never been more relevant. If your assets aren't connected to the internet, AI-powered web attacks can't reach them. That's not a complete solution to every threat, but it eliminates a significant portion of the attack surface that's currently being targeted.
The Uncomfortable Reality
We are past the point of treating AI-enabled cyberattacks as a future concern. They're happening now, at scale, and the tools enabling them are becoming cheaper and more capable every quarter. The security industry will adapt, it always does, but adaptation takes time and the gap between what attackers can do today and what most organizations are prepared to stop is wider than most security leaders publicly acknowledge.
Generative AI was supposed to be a productivity multiplier for legitimate work. It is that. It's also a force multiplier for everyone trying to break into systems, steal funds, and compromise infrastructure. Treating these two realities as separate conversations is no longer possible. The technology doesn't care which side of the law you're on, and neither do the models.
