Teams need runtime prevention that stops AI-based malicious threats before they execute

COMMENTARY: Open source software was born from a noble idea: transparency would lead to better software. Public codebases and shared vulnerability disclosures were meant to strengthen the collective, making it easier for developers to fix bugs and for defenders to harden systems.

However, in 2025, that same transparency has become a double-edged sword.

It’s time to face a hard truth: the rise of AI-speed attacks means the age of reactive security has ended. In this new era of always-present zero-day vulnerabilities and AI-crafted exploits, only prevention-first, secure-by-design strategies can effectively protect business-critical computing environments at scale, especially as AI shrinks the exploit window to near-zero and expands the scope of exploitable code far beyond what human teams could ever do.

[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]

Earlier this year, security researcher Sean Heelan published a blog that quietly shook the cybersecurity world. Using a custom AI model built on open-source bug data and fuzzing strategies, Heelan discovered CVE-2025-37899, a remote, unauthenticated, zero-day vulnerability in the Linux kernel’s SMB implementation.

An AI model identified a kernel-level remote code execution (RCE) vulnerability in one of the most widely deployed software components globally. No elite nation-state team. No sophisticated supply-chain backdoor. Just a well-trained model, public data, and a little bit of inquisitive hacking time.

This isn’t an outlier:

Vulnhuntr, an AI fuzzing system developed in 2024, uncovered more than a dozen exploitable bugs in Python’s core libraries, including RCE, LFI, and XSS vulnerabilities in projects like gpt_academic, ComfyUI, and Ragflow, none of which had CVEs at the time.
GitHub’s CodeQL team used LLMs to automatically model APIs and discovered a new CVE (CVE-2023-35947) in the Java/Gradle build toolchain which enabled an adversary to overwrite critical system files, exposing a opportunities to establish a foothold in the environment or deposit malware.
Google’s Big Sleep project, using LLM-driven static analysis, uncovered a previously unknown, exploitable memory flaw in SQLite—a database engine used in mobile and IoT environments—by identifying a stack buffer underflow that traditional fuzzing missed.

AI-assisted vulnerability discovery only stands as part of the story. Recent research from UC Berkeley’s RDI group, led by Dawn Song, demonstrates that today’s frontier AI models, including GPT-4 and Claude, cannot only identify vulnerabilities, but also generate fully functional proof-of-concept (PoC) exploits.

In controlled benchmarks, these models successfully reproduced working exploits for real-world memory safety CVEs, demonstrating that the entire attack chain, from bug to weaponization, can now be generated automatically by machines.

This compression of discovery-to-exploit timelines is unprecedented, and irreversible. Security teams can’t race against machines that don’t need rest, don’t get distracted, and improve with every cycle.

While this study focused on low-level C/C++ vulnerabilities, higher-level bugs in Python or Java—such as deserialization attacks or path traversal flaws — are likely even more vulnerable to LLM automation due to their simpler semantics and broader training exposure.

Why scan/patch and detect/respond can’t keep up

Traditional security was built on two key assumptions: that teams would identify vulnerabilities (so we could patch them), and that we could detect signs of exploitation (so we could respond). Both assumptions are now breaking down, and AI has accelerated their collapse.

Scan/patch timelines have already hit a floor. Most orgs patch on multi-week or multi-month cycles, with little room to move faster. Meanwhile, AI has effectively driven time-to-weaponization toward zero. It’s now possible to generate Proof-of-Concept (PoC) exploits minutes after a bug gets discovered, sometimes as part of the same AI process.

Furthermore, as AI builds the ability to analyze more code with ease, the attack surface continues to expand. That means the attacker’s advantage is no longer just stealth. It’s raw speed and an unprecedented scope of the attack surface.

Worse, scan/patch doesn’t apply when it's an unknown vulnerability. AI has sped-up known exploits while also discovering zero-days at scale. These vulnerabilities may never get published, never receive CVEs, and never show up in a scanner.

Detect/respond fares no better. AI-generated attacks increasingly resemble normal behavior, subtly hijacking app logic or runtime flows. With no clear indicators, even the best SIEMs and EDRs are often blind to execution paths that look “valid” on the surface.

Together, these shifts—instant weaponization and an ever-expanding unknown attack surface—mean defenders are chasing threats in a game that’s accelerating beyond human scale.

These vulnerabilities aren’t just theoretical—they lead to real-world, high-severity outcomes:

Library-level RCE in Python or Java allows lateral movement across microservices.
Container registry exposures make it easy to poison supply chains and tamper with AI models embedded in containers.
Kernel-level privilege escalation means an attacker owns the entire container host.
Cloud service misconfigurations discovered by AI allow privilege escalation across IAM roles in AWS, GCP, or Azure.

The cost isn’t just downtime. It’s reputational damage, customer churn, regulatory scrutiny, and in some sectors, national security risk.

If attackers can discover and weaponize vulnerabilities in minutes using AI, defenders can no longer rely on heuristic or signature-based detection rules, nor lengthy patch cycles. It’s not possible to scan for what isn't published or respond to what the team can't see, and that’s why the only effective defense against AI-speed threats is runtime prevention that stops malicious behavior before it executes. Because in many cases -- and what’s likely the common case going forward -- the attacker has already executed from within.

Ashar Aziz, co-founder, board member, BlueRock Security

SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.