The Bug-Finding Engines Were Never Asked to Slow Down
Seven months after machine-speed bug-finding arrived, every institution on record has adapted to it as permanent weather — and none has proposed slowing the engines down.
Daniel Stenberg maintains curl, the free data-transfer tool running across an enormous share of the internet's machines. For years he ran the standard arrangement for software like his — a bug bounty, paying outsiders who found genuine flaws in his code. Then the reports changed. The share of submissions that turned out to be real vulnerabilities fell from 15 percent to under 5 percent as machine-drafted filings poured in, so Stenberg ended the rewards [1]. In July he stopped accepting reports altogether. On HackerOne, the platform that runs such programs, report volume stayed double the prior year even after report quality recovered [1]. The smallest institution in this story had exactly one move available, and it took it: it closed its own door. What the rest of the world did on the other side of that door is the record of the past seven months. In April, two labs shipped engines that find vulnerabilities at machine speed. Anthropic kept its model inside a circle of roughly 50 vetted partners — Microsoft, Google, and JPMorgan among them — while OpenAI handed its version to thousands. Anthropic argued the speed was the point.
The work of defending the world’s cyber infrastructure might take years; frontier AI capabilities are likely to advance substantially over just the next few months. For cyber defenders to come out ahead, we need to act now. — Anthropic
The arithmetic that followed required no attacker. By May, the partners in Anthropic's circle had logged more than 10,000 high- or critical-severity findings [2]. By July, the industry's total of known vulnerabilities had doubled past 10,000 [3]. In October, the dashboard showed more than 6,000 new findings across 591 projects — and 516 of them patched [4]. Nearly nine in ten organizations still run exploitable flaws in production [4]. Microsoft's chief executive described the mechanism in plain terms this week.
Y2K was one bug, and one day. — Satya Nadella
Days before he said it, the mechanism had become visible at his own company. A latent bug left in a summer update took more than 80 percent of Microsoft's servers offline and disconnected millions of users for about a day [5]. The striking thing about every response since is that none treats the speed as something to change. Each treats it as weather, and builds accordingly. Insurers are pricing the condition. Allianz now lists AI as its corporate clients' second-biggest risk, and says the losses already arriving are "the tip of the iceberg" because they are being absorbed into existing cyber, liability, and directors-and-officers policies rather than counted separately [6]. The same insurer credits AI detection tools with materially lower breach costs — the honest wrinkle in a hazard story, since the technology is both the fastest-rising risk and a working defense [6]. One cyber-insurance executive set the industry's condition for staying insurable in five words.
Insurability will follow governability. — Jay Bavisi
Vendors are selling speed back as a product. Google's answer to the flood is a recommendation that enterprises modernize their triage and patch-management plans — move faster, not find less [3]. TrendAI feeds the engine's findings into "virtual patching," a stopgap that blocks an attack without fixing the underlying code [7]. Cisco has taken the same discovery model and resold it as a red-teaming service — hiring the engine to attack a company's own systems on purpose, to find the holes before an adversary does. Governments are doing the same arithmetic. The NSA is using the Anthropic model even as the Pentagon formally designates Anthropic a supply-chain risk — the same tool, flagged and adopted in the same breath. The Treasury secretary and the Federal Reserve chair summoned Wall Street executives over AI cyber threats, and OpenAI briefs federal agencies and the Five Eyes allies. After November's Microsoft outage, India demanded that the cloud giants be required to build self-reliant backup systems on Indian soil [8]. At the farthest edge, the adjustment has reached households. A "digital prepper" movement is stockpiling Faraday cages, offline copies of Wikipedia, and mesh radios against the possibility that the network itself fails [9]. The state is institutionalizing the instinct: Britain launched a "Be Ready" campaign in October, European leaders issued 72-hour self-sufficiency guidance this month, and the American vice president told citizens to be prepared for everything. None of this adds up to restraint. Across insurers, vendors, agencies, and the labs themselves, nothing on the record proposes capping, pausing, licensing, or throttling the engines that find the bugs. The only "restraint" ever tabled came from the labs in late September, and it was a call for independent testing before public release [10]. That checks who inspects the machines, not how fast they run. Critics and former employees read the move as positioning — shaping the rules and raising the bar against smaller competitors — and the White House dismissed the existential framing as a hoax [10]. So the year's one recorded act of refusal turns out to have been aimed at the reports, not the engines. Daniel Stenberg closed his own door. The insurers, vendors, agencies, and labs that came after him left theirs open and set about living with the speed — and in seven months of adaptation, not one of them has put a proposal on the record to slow the engines down.
- 1. AI-Generated Slop Floods Bug Bounty Programs With False Reports
- 2. Anthropic AI Tool Finds 10,000 Critical Vulnerabilities
- 3. Google and Gatekeeper Warn of AI-Driven Security Risks
- 4. AI-Driven Tooling Increases Security Dashboard Fatigue for Teams
- 5. Satya Nadella Warns AI-Driven Patching Increases Global Outage Risks
- 6. Allianz Warns AI Insurance Losses Are Tip of Iceberg
- 7. TrendAI Deploys Anthropic Claude Model to Automate Vulnerability Research
- 8. Microsoft Global Outage Sparks Warnings Over Systemic Tech Risks
- 9. Digital Preppers Prepare for AI and Network Collapse
- 10. OpenAI and Anthropic Call for AI Existential Risk Regulation