Swedish developer Daniel Stenberg, who has maintained the cURL library—a data transfer foundation for billions of devices, from routers to car firmware—for nearly three decades, closed the bug bounty program on HackerOne in January 2026. The reason: he spent most of his working time analyzing vulnerability reports generated by neural networks. The reports looked convincing, but contained no real flaws. Following him, Nextcloud made a similar decision in April. This is not a coincidence, but a systemic crisis of the model that has underpinned cybersecurity for decades.

AI Slop: When Quantity Kills Quality

The bug bounty economy was built on a simple principle: finding vulnerabilities is a rare and expensive skill, so the flow of submissions is naturally limited by the number of qualified specialists. Advanced AI models removed this limitation in a matter of months. John Morello, CTO of Minimus, cites a figure that explains everything: the share of genuine submissions has fallen from 15% to less than 5%. The rest is plausibly formatted "slop" that a neural network generates faster than an analyst can reject it.

Sophos CISO Ross McKerchar divides current market participants into three types: newcomers who run code through AI and submit the result without verification; experienced researchers who are themselves misled by AI agents; and entire teams that have built a pipeline for mass submissions from neural networks, designed for volume rather than accuracy. Platforms like Bugcrowd have already introduced mandatory identity verification, CAPTCHA, and strict blocks—up to a 30-day suspension for 10 rejected reports in a row. But these are half-measures.

Anthropic Experiment: Detection Speed Outpaces Fixing Speed

The industry experienced a real shock after Anthropic's experiment. The Claude Mythos 5 model, capable of autonomously finding zero-day vulnerabilities and writing working exploits for them, scanned over a thousand open-source projects. The result: 23,019 potential issues, of which 6,202 were priority or critical. Independent verification confirmed over 90% of the findings as real vulnerabilities. But less than 1% were fixed.

Among the gems: a 27-year-old vulnerability in OpenBSD allowing remote crash of any server; a 16-year-old one in FFmpeg that five million automated tests failed to detect; and a chain in the Linux kernel granting full device control. The problem is not that the holes exist—they always have. The problem is that the speed of their detection has, for the first time in history, outpaced the speed at which teams can verify, prioritize, and patch them.

Neither Claude Mythos 5 nor its counterpart from OpenAI—GPT-5.5-Cyber—have received a public release. Before releasing a tool that finds vulnerabilities faster than any human, the industry needs to close at least some of the holes that the neural network has already found in test mode.

The Market Shifts to Exclusivity

AI slop has collapsed the old economy: mass superficial reports have devalued simple vulnerabilities. Google reduced base rewards for memory errors in Chrome to $500 and canceled bonuses for simple bugs. In response, giants are forced to pay premiums for complex, rare, and confirmed exploits. Apple raised the ceiling for hidden access to the iOS kernel to $2 million, and considering beta versions and Lockdown Mode bypass, up to $5 million. Microsoft launched the Zero Day Quest for $5 million, and Google increased the reward for a zero-click exploit with persistence in the Titan M chip to $1.5 million.

This is not charity, but a forced defense against the shadow market, where private brokers offer up to $7 million for a full attack chain for current versions of iOS or Android. The market hunts for a small group of engineers whose skill in manually assembling complex exploits is worth millions against the backdrop of cheap automation.

Regulators and the Future: This Is Just the Beginning

On September 11, 2026, the EU's Cyber Resilience Act comes into force, requiring companies selling network-connected products to report actively exploited vulnerabilities within 24 hours. The fine is up to €15 million or 2.5% of global annual revenue. The law requires maintaining a reporting channel, which is already overwhelmed by AI slop. If cURL could close its bug bounty relying on volunteers, corporations do not have that option.

The problem is exacerbated by the asymmetry between attack and defense. The window between vulnerability discovery and exploitation shrinks to minutes. Meanwhile, even cheap open-source models with 3.6–5.1 billion parameters, with proper context input, successfully reproduce findings at the level of CVE-2026-4747 from Anthropic's cases.

The crisis in cybersecurity is the first visible symptom of a more 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 these findings and turning them into closed vulnerabilities. Similar dynamics are already emerging in jurisprudence and content production: AI generates faster than specialists can verify.

My conclusion: We stand on the threshold of an era where the main resource becomes not the speed of detection, but the ability to verify and patch. Vendors and regulators have yet to answer the question of who will close this gap. But one thing is clear: the old bug bounty model is dead. It is being replaced by a market where only what has passed manual verification holds value.