AI found the bug before you did

For most of its 30-year life, cURL has been quietly reviewed by humans. Developers, security researchers, and static analysis tools have pored over its source code in rounds of audits. Then, in early 2026, something shifted. Daniel Stenberg, the creator and lead maintainer of cURL, posted a note that would have been unthinkable a year earlier: the AI slop bug reports had stopped. In their place, an ever-increasing stream of genuinely good security reports was arriving, almost all done with the help of AI, submitted at a frequency he had never seen before. Just three months into 2026, the cURL team had found and fixed more vulnerabilities than in each of the previous two years. Stenberg himself used AI to scan his own code and, with one click, it flagged over 100 bugs that had survived rounds of human review and traditional static analyzers. "Almost magical," he called it. This is not a story about one tool or one project. It is the story of a threshold being crossed. AI has become genuinely better than most humans at finding software bugs, and the implications ripple far beyond any single codebase.

The shift became real in early 2026

Stenberg's experience tracks with a broader pattern. The NPR article covering Anthropic's Project Glasswing quoted him directly: improvement in AI models' capabilities for bug-finding became noticeable in early 2026, following the release of new cutting-edge models in late 2025. The quality gap closed fast. Where 2024 and 2025 had been defined by a flood of low-quality, AI-generated "slop" reports that overwhelmed maintainers and ultimately led Stenberg to shut down cURL's bug bounty program entirely in January 2026, the reports arriving now were different. They were specific, accurate, and actionable. ZeroPath, an AI-powered static analysis tool, exemplified the change. Its scanner uncovered 170 verified issues in cURL alone, spanning HTTP/3, SMTP, IMAP, and TFTP, finding C footguns and logic bugs that no previous tool had caught. Joshua Rogers, a security researcher who tracked the results, noted that approximately 98% of the bugs recently reported and fixed in the cURL codebase were discovered by ZeroPath. That is a staggering number for a project that has been under continuous human scrutiny for three decades. Wordfence, which tracks vulnerability research in the WordPress ecosystem, reported a 453% increase in submission volume since October 2025, with AI-assisted vulnerability research overtaking non-AI-assisted reports by March 2026. The shift was not gradual. It was sudden.

Project Glasswing as a case study

If the cURL story represents the grassroots version of this shift, Anthropic's Project Glasswing is the institutional one. Announced in April 2026, Glasswing is a cybersecurity initiative built around Claude Mythos Preview, an unreleased frontier model that Anthropic explicitly decided was too dangerous to release publicly because of its cybersecurity capabilities. The numbers are striking. Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. Among the discoveries: a 27-year-old bug in OpenBSD, an operating system specifically designed for security, and a 16-year-old vulnerability in FFmpeg that automated testing tools had failed to detect despite running the affected code line five million times. Anthropic did not train Mythos specifically for cybersecurity. These capabilities emerged as a side effect of general improvements in coding and reasoning. That detail matters. It means this is not a specialized security product that can be contained to one niche. Every frontier model that gets better at understanding code will, as a byproduct, get better at breaking it. The coalition Anthropic assembled for Glasswing reads like a who's-who of tech: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, plus over 40 additional organizations that maintain critical software infrastructure. Anthropic committed up to $100 million in usage credits for the initiative. This is not a research experiment. It is an acknowledgment that the vulnerability landscape has fundamentally changed.

The asymmetry problem

Here is the uncomfortable math of software security: defenders need to find all the bugs. Attackers need to find one. This asymmetry has always existed, but AI amplifies it in both directions simultaneously. The same capability that lets Mythos Preview find thousands of vulnerabilities for defenders also means that equivalent models, when they inevitably proliferate, will find those same vulnerabilities for attackers. As one Forrester analyst noted, discovery is no longer the bottleneck. Interpretation, prioritization, and remediation are. Bradley Smith, Deputy CISO at BeyondTrust, put it bluntly in his response to the Glasswing announcement: "Those who are presenting it as giving the good guys a head start mischaracterizes where we actually are. The adversary already has AI working for them." His team had already observed AI-assisted tooling compressing the exploitation window for critical vulnerabilities to minutes, not weeks, and that was with current-generation tools, months before Mythos existed. The CNN report on Mythos captured the escalation dynamic perfectly: behind Mythos is the next OpenAI model, and the next Google Gemini, and a few months behind them are the open-source models. When open-weight models reach this capability threshold, which credible estimates put at months rather than years, the volume and sophistication of AI-driven attacks scales to a level most organizations are structurally unprepared for.

The bug bounty economy is already breaking

The economic effects are already visible. Stenberg ended cURL's bug bounty in January 2026 after six years, $86,000 paid out, and 78 confirmed vulnerabilities fixed. The program was drowning in garbage reports, and the team needed to stop the bleeding. The Internet Bug Bounty program followed suit, stopping monetary awards at the end of March. But the story took an unexpected turn. After the bounty ended and the financial incentive for slop disappeared, the quality of AI-assisted reports improved dramatically. HackerOne paused bug bounties for a different reason: with automated discovery producing such high volumes of legitimate findings, the bottleneck shifted entirely to remediation, which bounties do not fund. Forrester's analysis of Project Glasswing laid out the implications for the penetration testing industry directly: traditional pentests that run between $20,000 and $120,000, priced around the perceived scarcity of discovery expertise, face revenue erosion because Mythos Preview surfaced thousands of comparable vulnerabilities autonomously in weeks, without billable hours. Finding bugs is no longer the differentiator. For security researchers, this does not mean obsolescence, but it does mean the job description is changing fast. The value is shifting from "can you find the bug" to "can you understand the codebase, assess real-world exploitability, prioritize remediation, and guide organizations through fixing it." The mechanical work of scanning code for known patterns is being automated. The judgment work remains human, for now.

The irony: AI also generates buggy code

There is a rich irony in AI becoming the best bug finder while simultaneously being a prolific bug creator. The data on AI-generated code security is not encouraging. Veracode's 2025 research found that 45% of AI-generated code introduces OWASP vulnerabilities. CodeRabbit's analysis showed AI-produced code carries a 2.74x higher vulnerability rate than human-written code. Escape.tech scanned 5,600 applications built primarily with AI coding tools and found over 2,000 vulnerabilities, 400+ exposed secrets, and 175 instances of personally identifiable information sitting in the open. Stack Overflow's analysis of 470 GitHub repos found that AI created 1.7 times as many bugs as humans overall, with 1.3 to 1.7 times more critical and major issues. The New York Times described the result as "code overload": AI coding tools have enabled so much code to be written so quickly that companies cannot keep up with reviewing, testing, and securing it. The same technology that accelerates creation accelerates the creation of vulnerabilities. This is not a contradiction. It is the same capability expressing itself in two directions. AI is extraordinarily good at pattern-matching across code, which makes it excellent at both writing plausible code and finding flaws in code. The difference is that writing code is generative and error-prone by nature, while scanning code is analytical and benefits from exhaustive coverage. AI's strengths map better to the second task.

What this means for smaller teams

If you are an indie developer or running a small team, the practical question is: what do you do differently right now? First, integrate AI-powered security scanning into your workflow. Tools like ZeroPath, Snyk, and others are finding real bugs that traditional linters and SAST tools miss. The cost of not using them is rising fast as the attack surface grows. Second, assume your code has vulnerabilities that no human reviewer has caught. This was always true, but the gap between what AI can find and what human review catches has made it undeniable. If an AI scanner can flag over 100 bugs in cURL, a project maintained by one of the most careful developers in open source for 30 years, your codebase is not immune. Third, prioritize remediation infrastructure over discovery. The bottleneck is no longer finding the bug. It is fixing it, testing the fix, deploying it, and verifying the fix did not break something else. Invest in CI/CD pipelines, automated testing, and dependency management. Fourth, if you are using AI to write code, use AI to audit it too. Do not trust the output of one model without verification. The irony of AI-generated vulnerabilities being caught by AI-powered scanners is real, but it is also the most practical defense available right now.

The 30-year-old codebase and the new world

There is something poetic about cURL being at the center of this story. It is a tool that predates the modern internet as most people know it. It runs on billions of devices, from cars to medical equipment to every major operating system. It was written, maintained, and secured by humans for three decades. And now, in a matter of months, AI has found more bugs in it than years of human review. Not because the humans were careless, but because the scale of what AI can analyze, the patterns it can detect across millions of lines of code, the speed at which it can test hypotheses, these are capabilities that do not have a human equivalent. Stenberg's own words capture the shift: "LLMs have now bypassed human capability for bug finding." Coming from someone who has maintained one of the most widely deployed pieces of software on the planet, that is not hype. It is observation. The question going forward is not whether AI will dominate vulnerability discovery. That question is already answered. The question is whether defense can scale faster than offense. Project Glasswing is a bet that it can, if the right organizations have access to the right tools and the political will to act on what they find. The 90-day window that Glasswing's coalition has before equivalent capabilities proliferate more widely is not a comfortable margin. But the trajectory is clear, and it is, on balance, a net positive. Software that has been running with hidden flaws for decades is being secured. Vulnerabilities that survived millions of automated test runs are being caught. The defenders have a new tool, and it is the most powerful one they have ever had. The catch is that the attackers have it too.