Something historic happened 2 weeks ago…

A researcher used OpenAI’s o3 model—a powerful large language model (LLM)—to uncover a zero-day vulnerability in the Linux kernel. That’s not just a headline; it’s a turning point.

AI didn’t just assist—it discovered, explained, and proposed solutions for a complex security flaw in a core infrastructure component. And it did this without custom scripts or complex tools.

Let’s talk about what this really means for the future of cybersecurity—both the promise and the peril.

AI vs. the Linux Kernel

On May 20, 2025, a zero-day bug (CVE-2025-37899) was confirmed in the Linux kernel’s SMB component. It’s technical, but here’s the essence:

The bug, a “use-after-free” vulnerability, allows attackers to hijack memory that was already freed—enabling them to potentially run code with system-level privileges.

What’s groundbreaking is how it was found: a researcher ran the o3 model on about 12,000 lines of code. The AI didn't just find the bug. It:

• Identified a previous fix that would have failed

• Diagnosed a rare race condition across concurrent sessions

• Offered a full report that "felt human-written"

This wasn’t just an assist—it was a full-on discovery, traditionally done by experts using symbolic execution, fuzzing, or months of review.

Big Deal?

Until now, LLMs were known for chat interfaces, writing emails, or summarizing documents. This discovery shows they can operate like expert security researchers—able to understand timing-based flaws, complex memory issues, and logic bugs.

The key takeaway?
AI is becoming good at things we once thought only humans could do.

That opens the door to a new kind of cybersecurity research: AI-assisted discovery at scale.

But there’s a flip side.

The Double-Edged Sword: Defenders and Offenders

Just as AI is helping discover bugs, attackers are already using it to build malware, automate phishing, and manipulate trust in text, video, and voice.

Threat actors are using AI to:

• Jailbreak ethical safeguards (WormGPT, GhostGPT)

• Generate social engineering scams in multiple languages

• Create real-time deepfakes for impersonation fraud

• Write malicious code at a level that once required expert developers

This isn’t theory. FunkSec, a known ransomware group, claims 20% of their operations are now AI-powered. Nation-state groups, particularly from Iran and China, are reportedly leveraging LLMs in every phase of their attack cycle—from reconnaissance to command-and-control systems.

We are no longer in a world where you need a sophisticated hacker team to run an attack. All you need is access to the right model—and a prompt.

We’ll be right back.

Keep more of what you earn

Collective helps members keep more of what they earn — saving an average of $10,000 a year in taxes* — while taking countless hours of administrative work off their plates.

Your membership includes LLC and S Corp formation, payroll, monthly bookkeeping, quarterly tax estimates, annual business tax filing and more.

Get your first month free with code SPRINGFREE

_{*Based on the average 2022 tax savings of active Collective users with an S Corp tax election for the 2022 tax year}

So, What Does the Future Hold?

Opportunity:

• AI-Powered Defense Tools
  AI is revolutionizing how we detect threats—spotting behavioral anomalies, analyzing malware, and mapping attack patterns faster than any team could.

• Better Vulnerability Research
  As demonstrated by o3, LLMs can now reason about logic, concurrency, and memory issues. They’re like having 100 junior researchers that don’t sleep or get bored.

• Automated Threat Intel
  From scanning millions of phishing domains to flagging shady file names and automating rule creation for defense tools—AI is becoming an analyst's best friend.

Risk:

• Weaponized AI Models
  Malicious actors are customizing LLMs specifically for cyber crime. GhostGPT and FraudGPT are sold on Telegram with jailbreaks pre-baked in.

• Scale and Accessibility
  Open-source models mean even low-skilled attackers can launch sophisticated campaigns.

• Deepfake Explosion
  AI-generated voices and videos are already undermining basic trust in audio and visual media. From fake CEOs authorizing wire transfers to fabricated police officers coercing victims, it’s all happening now.

If you're curious about where this goes next, we’ll keep breaking it down. That’s what we’re here for.

This Week in AI

• Apple brings Apple Intelligence to the Iphone

• OpenAI hits $10B in ARR

• Amazon planning to invest $20B in data centres