Swedish developer Daniel Stenberg, who has maintained the cURL library — the data transfer foundation for billions of devices, from routers to car firmware — for nearly thirty years, shut down the bug bounty program on HackerOne in January 2026. The reason: for the past two years, most of his time was spent analyzing vulnerability reports generated by neural networks. The reports looked convincing, but contained no real vulnerabilities. Following cURL, Nextcloud made a similar decision in April of the same year.

This is not a coincidence, but a symptom of a systemic crisis. The bug bounty model, which relied for decades on a shortage of qualified personnel, has collapsed under the onslaught of AI slop. John Morello, CTO of Minimus, provides telling statistics: the share of genuine submissions has dropped from 15% to less than 5%. The rest are plausible but empty reports, which a neural network generates faster than an analyst can reject them.

Three types of bug "hunters"

Ross McKerchar, Chief Information Security Officer at Sophos, divides modern submitters into three categories:

  • Novices who run code through AI and submit the result without checking whether it's a real vulnerability or a hallucination.
  • Experienced researchers who are misled by the AI agents themselves: the neural network convincingly reports a non-existent problem, and the person trusts it.
  • Teams that have built a pipeline of AI agents for mass submission, designed for volume rather than accuracy.

Major platforms are trying to fight back. HackerOne launched the Hai Triage system to filter AI-generated submissions. In March 2026, Bugcrowd introduced mandatory identity verification, dynamic rate limiting, and CAPTCHA, along with strict penalties: from a 30-day suspension for 10 consecutive rejected reports to a permanent ban for automated spam.

Anthropic's experiment: detection speed vs. remediation speed

The scale of the problem became evident after Anthropic's experiment with the Claude Mythos 5 model. This LLM can autonomously find zero-day vulnerabilities and write working exploits for them. As part of Project Glasswing, the model scanned over a thousand open-source projects, identifying 23,019 potential issues, of which 6,202 were classified as high-priority or critical. An independent verification of a sample of 1,752 findings confirmed over 90% as real vulnerabilities. However, less than 1% of the total findings were fixed by specialists.

This is a historic turning point: the speed of vulnerability detection has, for the first time, outpaced the speed of verification and remediation. Mythos 5 found a 27-year-old vulnerability in OpenBSD, a 16-year-old one in FFmpeg (which 5 million automated tests failed to detect), and a chain in the Linux kernel granting full device control.

The market shifts toward exclusivity

AI slop has shattered the old bug bounty economy: mass superficial reports have devalued simple vulnerabilities. Google lowered base rewards for memory bugs in Chrome to $500 and canceled bonuses for simple bugs. At the same time, a polished Proof-of-Concept commands unprecedented sums: Apple raised the ceiling for iOS kernel access to $2 million (with the potential to reach $5 million for reports in beta versions), Microsoft launched the Zero Day Quest competition with a $5 million prize pool, and Google increased the reward for a zero-click exploit with persistence in the Titan M chip to $1.5 million. This is a forced defense against the shadow market, where a full attack chain can fetch up to $7 million.

The crisis is just beginning

The situation is exacerbated by regulatory pressure. On September 11, 2026, the EU's Cyber Resilience Act comes into effect, requiring companies to report actively exploited vulnerabilities within 24 hours, under threat of fines up to €15 million or 2.5% of global annual revenue. The law requires a functioning channel for receiving reports — the very channel that is being overwhelmed by AI slop.

The problem is not limited to top-tier LLMs. Company Aisle demonstrated that even small, cheap open-source models (with 3.6–5.1 billion active parameters) successfully detect key vulnerabilities like CVE-2026-4747 when given the right context.

My analysis: The cybersecurity industry was designed for a world where the main problem was detection. That world is over. The new bottleneck is verification, prioritization, and patching. AI acts as a producer of findings on a scale unimaginable just three years ago, while humans remain the only link capable of verifying those findings and turning them into closed vulnerabilities. The question is not whether we can find bugs faster, but whether we can fix them fast enough to avoid drowning in our own code.