A tectonic shift has occurred in the world of cybersecurity. Swedish developer Daniel Stenberg, who has maintained the cURL library — critical infrastructure for billions of devices — for nearly three decades, shut down the bug bounty program on HackerOne in January 2026. Nextcloud followed suit with a similar decision in April of the same year. The reason is not a lack of vulnerabilities, but an excess of them, generated by artificial intelligence.
The model where a researcher finds a bug, reports it, and receives a reward has collapsed under the onslaught of AI slop. Minimus CTO John Morello provides sobering statistics: the share of genuine submissions has dropped from 15% to less than 5%. The rest are convincing but false reports generated by neural networks faster than human analysts can reject them.
Three Types of Claimants in the New Reality
Sophos Chief Information Security Officer Ross McKerchar identified three categories of bug hunters in the AI era:
- Novices — run code through AI and submit the result without verifying its accuracy;
- Experienced researchers — trust AI agents that convincingly report non-existent problems;
- Teams — build mass submission pipelines focused on volume rather than accuracy.
HackerOne has already launched an AI-based Hai Triage system to filter generated submissions, while Bugcrowd has introduced mandatory identity verification, CAPTCHA, and strict blocking measures. But these are half-measures. The decision by cURL and Nextcloud is a rational response to an economy that has stopped working.
Anthropic Experiment: Detection Speed Outpaces Fixing Speed
The scale of the problem became evident after testing Claude Mythos 5 — a model capable of autonomously finding zero-day vulnerabilities and writing working exploits. As part of Project Glasswing, the model scanned over a thousand open-source projects, identifying 23,019 potential issues. Of the 1,752 verified findings, over 90% were confirmed as real vulnerabilities. However, less than 1% were fixed.
This is a catastrophic gap: detection speed has, for the first time in history, outpaced verification and patching speed. Mythos 5 found a 27-year-old vulnerability in OpenBSD, a 16-year-old one in FFmpeg (which 5 million automated tests missed), and a chain in the Linux kernel granting full device control.
Vulnerability Market: From Mass to Exclusive
AI slop has devalued simple bugs. Google reduced base rewards for memory errors in Chrome to $500, while Apple raised the ceiling for hidden access to the iOS kernel to $2 million. Microsoft launched the Zero Day Quest competition with a $5 million prize pool. But this is not charity — it's a fight against the shadow market, where private brokers pay up to $7 million for a full attack chain. Everyone is hunting for a small group of engineers whose skills are worth millions against a backdrop of cheap automation.
Regulatory Zugzwang
On September 11, 2026, the EU's Cyber Resilience Act comes into effect, requiring companies to report actively exploited vulnerabilities within 24 hours. The fine is up to €15 million or 2.5% of global revenue. The law requires a functioning bug reception channel, which is being overwhelmed by AI slop. While cURL could close its program relying on volunteers, corporations selling software in the EU have no such option.
The problem is compounded by the fact that even cheap AI models can replicate Anthropic's findings. Company Aisle demonstrated that small models with 3.6–5.1 billion parameters, when properly configured, successfully find critical vulnerabilities.
My analysis: The crisis in cybersecurity is just the first symptom of a fundamental shift. AI has become a producer of findings on a scale unimaginable three years ago, while humans remain the only link capable of verifying them. Similar dynamics are already visible in jurisprudence and content production. The era when the main problem was vulnerability detection is over. The new bottleneck is verification, prioritization, and patching. And neither vendors nor regulators know who will close this gap. Investors should look at startups automating these processes, rather than bug hunting.