The wave of automated intrusions sweeping through the technology sector over the past months is not an isolated incident — it is the symptom of a paradigm shift that global cybersecurity can barely articulate. What happened two years ago only in state intelligence manuals now runs on conventional workstations at a cost of cents per execution.
The central report on the topic, published on The Verge's Decoder podcast last week, features Alex Stamos, chief security officer at Cognition and researcher at Stanford University's Cyber Policy Center, describing a scenario he calls a "dark age." The metaphor is not gratuitous: Stamos compares the current situation to the early internet era of the 1990s, when small underground groups discovered vulnerabilities at scale — but with one crucial difference. Back then, networks were primarily academic and isolated; today, the same tools expose the connected global commerce.
What changed technically? According to Stamos, so-called "open-weight models" — artificial intelligence models whose weights are publicly accessible and can be run locally — have removed the barrier that previously limited automated attacks to commercial AI services with safety guardrails and law-enforcement-traceable logs. A quantized, "obliterated" GLM-5.3 model (one stripped of its built-in safety filters) runs on a Mac Studio with consumer hardware and can run offensive tooling independently.
But the most revealing concept Stamos presented is "kill chain inversion." Traditionally, threat actors spent weeks or months conducting manual vulnerability research, writing exploit code, and staging human operators to deploy payloads against selected targets. The operation was essentially a stockpile process: research, create, store, then fire. The new flow inverts this logic. Automated software engines inspect target infrastructure in real time, detect vulnerabilities, and generate operational exploit code specific to that environment, without prior human intervention. The speed of discovery eliminates the defensive response window that security teams traditionally used to patch known weaknesses.
A concrete case illustrates the scale: the ARTEX framework, a Chinese open-source tool with an English translation available on GitHub. When connected to an uncensored open-weight model back-end, such as the obliterated GLM-5.3, the system scans target networks, maps unpatched weaknesses, and writes custom payloads. Campaigns with this architecture began with three South Korean banks and, within a week, hit approximately 200 additional organizations.
What makes the problem even more complex is the developer tooling supply chain. Recent attacks exploited AI infrastructure providers — the Fireworks AI case, where API credentials were exfiltrated from customer evaluation jobs, and then used to access the Cursor platform, where a malicious plugin was distributed to up to 5,000 corporate environments. Each of those companies became, potentially, a secondary entry point.
On the defensive side, the economic problem is equally devastating. Stamos points out that the marginal cost of tokens for attackers using open models on commodity hardware is close to zero. For the defender, however, the equation is radically different: security teams deal with thousands of simultaneous vulnerabilities, each requiring manual triage, prioritization, and remediation. Bug bounties, security conferences, and responsible disclosure norms — mechanisms that maintained some equilibrium for decades — do not scale at the speed automated agents operate.
Benchmark data corroborates the urgency. Vals AI tests showed that DeepSeek V4.1 Flash (open-weight model) completed about 73% of objectives in its CyberOffense ranges against 81% for GPT-6 Sol (OpenAI's restricted model), costing under $3 per execution versus $22 for the proprietary model. The performance gap narrows further when you consider that open models need more tokens to achieve significant progress — short evaluations systematically underestimate their capability. The UK AI Security Institute estimated that the lag between open and closed models fell to four to seven months, down from six to ten months in 2025.
The sector's response has begun, but it remains fragmented. Mozilla now ships weekly Firefox updates to shrink the patch gap. Japan approved a ¥10 billion emergency cyber package. OpenAI classified its GPT-6 Cyber as "Critical" for security and limits it to approved Daybreak Red customers. Cognition, the company where Stamos works, has dispatched teams to help compromised businesses. But none of these movements scale at the necessary speed.
The question is whether digital defense infrastructure will be able to adapt before attackers reach full operational maturity — or if we are headed for years of sustained operational friction until organizations completely reimagine their network architectures, isolate sensitive development environments, and deploy autonomous agents capable of patching compromised code at machine speed.
Sources: The Verge, MLQ News, SafeDep
✓ Independent sources cross-checked and verified before publishing
