AI Inverts the Cyber Kill Chain, Pushing Tech into a "Dark Age" of Breaches

The multibillion-dollar cybersecurity industry built over the past two decades is facing an existential crisis. According to Alex Stamos, chief security officer at Cognition and lecturer at the Stanford University Cyber Policy Center, the rise of open-weight artificial intelligence models and automated attack tools has pushed the digital world into a security "dark age."

Speaking on The Verge's Decoder podcast, Stamos argued that the economics and operational architecture of cyber defense have been completely scrambled. While traditional computer security relied on a slow, human-driven process of vulnerability discovery followed by patch deployment, AI systems have automated almost the entire "kill chain" in real time.

An Inverted Cyber Kill Chain

Historically, executing high-level cyberattacks required significant resources. Elite researchers spent months discovering vulnerabilities and writing exploit code, which operational teams or state-backed groups then deployed against targets. That hierarchy—ranging from amateur "script kiddies" to a select few hundred top-tier vulnerability researchers—is rapidly collapsing.

"What has happened over the last month is we’ve had a couple of these big bugs, PHP, Node and such. What’s been less discussed is a huge number of companies have been breached," Stamos explained. Attackers are now chaining exploits across vendors: an intrusion at AI service provider Fireworks compromised API keys that opened access into Cursor, which was subsequently used to push a malicious plugin affecting 5,000 client companies.

Stamos highlighted open-source tools such as ARTEX, a Chinese framework available on GitHub. When paired with open-weight models like an obliterated GLM-5.3 running on local hardware—such as Apple Mac Studio workstations or consumer gaming rigs—these systems can autonomously scan targets, discover vulnerabilities, and write stable exploits simultaneously. Recent campaigns utilizing these capabilities compromised three South Korean banks before hitting approximately 200 additional organizations within a week.

The Asymmetric Economics of AI Exploits

The disruption is fundamentally economic. For attackers, the marginal cost per token is plummeting toward zero, enabling small teams or individual actors to run continuous autonomous attack agents. Meanwhile, defensive costs remain unsustainably high. Stamos noted that running a single investigation using a leading security platform's frontier model can cost up to $900 per incident—an impossible expense when defending against relentless, automated probes.

Furthermore, state-backed actors continue to exploit specialized vectors alongside automation. The recent breach of Binance—described by Stamos as the largest heist in human history—was carried out by North Korean teams leveraging sophisticated human intelligence and social engineering rather than automated script tools. Meanwhile, smaller Chinese contractors in the "cottage access industry" breach targets to sell entry points to larger domestic enterprises.

Beyond Patchwork: The Case for Rewriting Infrastructure

Stamos warned that the tech industry cannot simply patch its way out of this environment. Frontier models like Mythos and Astra Cyber are uncovering thousands of previously hidden vulnerabilities across foundational open-source codebases, generating continuous cycles of emergency updates.

Rather than perpetually fixing legacy C code, Stamos urged cloud giants to take structural action:

• Autonomous remediation: Organizations must move toward automated bug fixing, where AI agents identify flaws, write extensive test suites, and refactor code directly into memory-safe and type-safe languages.
• Rewriting core infrastructure: Stamos called on Google and Amazon to collaborate on forking and rewriting the Linux kernel in Rust for enterprise cloud infrastructure. Eliminating unneeded legacy drivers and moving to memory-safe architectures would mitigate recurring kernel-level patch cycles that disrupt global server fleets.
• Autonomous SOCs with deterministic limits: Security operations centers must adopt tiered AI agents to parse petabytes of network logs. However, Stamos cautioned that defensive agents must have hard deterministic constraints—such as capping the number of accounts an AI can disable in identity systems like Okta—to avoid taking down entire companies during false positives.

Proactive Defense in the Next 90 Days

For enterprise leaders, passive risk management is no longer viable. Stamos warned that every enterprise must deploy AI red-teaming tools against their own infrastructure within the next 90 days or face inevitable compromise.

"You are going to have to take that risk to actively red team yourself, to front run the bad guys who are doing that right now," Stamos said. While defensive capabilities will eventually mature through autonomous remediation and memory-safe design, the immediate future demands that organizations adapt rapidly to survive the escalating speed of automated threats.