The bug bounty model, which has served as the foundation of cybersecurity for decades, is experiencing a systemic crisis. Swedish developer Daniel Stenberg, who has maintained the cURL library for over 30 years, recently announced the closure of the reward program on HackerOne. The reason is an avalanche of vulnerability reports generated by neural networks that look convincing but contain no real threats. Following his lead, the developers of Nextcloud made a similar decision.

The bug bounty economy was built on a scarcity of skilled researchers. Artificial intelligence removed this constraint in a matter of months. According to estimates by John Morello, CTO of Minimus, the share of genuine submissions has dropped from 15% to less than 5%. The rest is plausible but empty AI noise ("slop"), generated faster than analysts can reject it.

Three Types of Claimants in the New Reality

Sophos Chief Information Security Officer Ross McKerchar identifies three categories of market participants: newcomers who blindly trust AI models; experienced researchers misled by neural network hallucinations; and organized teams that have turned AI agents into a conveyor belt for mass submissions, focused on volume rather than accuracy.

The largest platforms are responding harshly. HackerOne has launched the Hai Triage filtering system, while Bugcrowd has introduced mandatory identity verification, CAPTCHA, and account blocking after 10 consecutive rejected reports. The decisions by cURL and Nextcloud are not a capitulation but a rational response to an economy that has ceased to function.

Anthropic Experiment: Detection Speed Outpaces Fixing Speed

The scale of the problem became evident after the launch of Anthropic's Claude Mythos 5. The model, capable of autonomously finding zero-day vulnerabilities and writing working exploits, scanned over 1,000 open-source projects. The result: 23,019 potential issues, of which 6,202 were deemed priority or critical. Independent verification confirmed over 90% of the findings as real vulnerabilities. Yet less than 1% have been fixed.

Three findings are particularly telling: a 27-year-old vulnerability in OpenBSD allowing remote server crashes; a 16-year-old one in FFmpeg missed by 5 million automated tests; and a chain in the Linux kernel granting full device control. Neither Claude Mythos 5 nor the similar GPT-5.5-Cyber from OpenAI received a public release — the industry cannot keep up with patching the holes they find in test mode.

The Market Shifts Toward Exclusivity

AI slop has crashed the price for simple vulnerabilities. Google has lowered base rewards for memory bugs in Chrome to $500, while Apple, Microsoft, and Google are forced to pay unprecedented premiums for complex, confirmed exploits. The reward ceiling for hidden access to the iOS kernel has reached $2 million, and for a zero-click exploit with persistence in the Titan M chip — $1.5 million. This is not charity but a forced defense against a shadow market offering up to $7 million for a full attack chain.

All market participants are now hunting for a small group of engineers whose skill in manually assembling complex exploits is worth millions amid cheap automation.

Regulators Intensify Pressure

On September 11, 2026, the EU's Cyber Resilience Act comes into force, requiring companies selling connected devices 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 channel for receiving submissions, which is already drowning in AI noise. While cURL could close its program relying on volunteers, corporations do not have that option.

The problem is compounded by the fact that even old and cheap models can reproduce Anthropic's findings. An experiment by Aisle showed that small neural networks with 3.6–5.1 billion parameters, given the right context, successfully detect key vulnerabilities.

The cybersecurity crisis is the first symptom of a fundamental shift. AI acts as a producer of findings on a scale unimaginable three years ago, while humans remain the only link capable of verifying and fixing them. Similar dynamics are already emerging in law and content production. Across all fields, a division of labor is taking shape: AI produces the product, and the specialist bears responsibility for quality.

My analysis: The industry has found itself trapped by its own progress. The speed of vulnerability detection has outpaced the speed of fixing them, and this gap will only grow. The question is not who will find more holes, but who will be able to close them. Neither vendors nor regulators have an answer yet.