cURL Ends Bug Bounty Over AI-Generated Fake Vulnerability Reports

AI-Generated Fake Reports Overwhelm cURL's Bug Bounty

Daniel Stenberg, creator of cURL, detailed how the project's HackerOne bug bounty program — active since April 2019 and having paid out over $85,000 — became flooded with AI-generated vulnerability reports. In a review of 500 submissions, only 10–15% were valid; the rest were lengthy, confident-sounding but entirely fabricated reports produced by large language models.

A concrete example: one report claimed an HTTP/3 "stream dependency loop vulnerability" rated as "extremely severe, life-threatening." It included a GDB session and register dumps, yet the referenced functions did not exist in cURL's codebase. Everything was fabricated.

Why the Bounty Program Was Stopped

Stenberg identified the bounty structure as a key driver: rewards up to $10,000 for critical bugs and $500 for low-severity ones incentivized submitters to run AI tools, paste results, mark them "critical," and hope for a payout — with almost no verification. By late 2025, roughly 1 in 20–30 reports was AI-generated junk, turning triage into a "horror show" that exhausted the seven-person security team's time, energy, and morale.

Stenberg argued that if maintainers become desensitized by noise, real vulnerabilities will slip through, endangering the broader software supply chain. To remove this perverse incentive, the team formally closed the bug bounty program.

Shift to Public Disclosure for Low/Medium Severity Issues

Starting October 26, 2026, low- and medium-severity reports will be handled in cURL's public GitHub repository. Stenberg noted that over 90% of past submissions were quickly assessed as below "high" severity. Public handling does not reduce per-report effort but distributes the workload and makes lessons and documentation more discoverable. The trade-off: attackers gain more time to exploit before a fix ships, but Stenberg considers this acceptable for non-critical issues.

AI as a Tool for Finding Real Vulnerabilities

Despite the abuse, Stenberg emphasized that AI analysis tools, used by experienced engineers, have uncovered "100+" genuine vulnerabilities that survived aggressive compiler flags, fuzzing, traditional static analysis, and multiple human audits. These tools reason across protocols, specifications, and third-party libraries in a "near-magical" way. For instance, an AI flagged an invalid byte in a Telnet implementation per a 2012 spec that "nobody has looked at since 2012." They also catch inconsistencies between function comments and implementations, revealing subtle logic errors.

The project employs three different AI code-review bots that run at 2 AM when no human reviewers are active. They find missing tests, incorrect assumptions about external library behavior, and other issues — though they do not replace the test suite.

Skepticism Toward AI-Generated Production Code

Stenberg remains deeply skeptical of using AI to write production code. He does not use AI coding tools himself and believes no cURL team member relies on them for serious development. AI-suggested code fixes are "never good enough" to accept blindly; he likens them to an "enthusiastic junior programmer" whose suggestions require careful vetting and full testing.

Legally, AI-generated contributions do not alter cURL's risk model: the project has always trusted that contributors have the right to submit their code, regardless of whether it was hand-written, copied, AI-generated, or sourced from Stack Overflow.

Broader Open Source Context: FFmpeg vs. Google

Stenberg situated cURL's experience in a wider pattern. He cited the FFmpeg–Google dispute, where a well-resourced company used AI and large security teams to find numerous real vulnerabilities, then pressured a small volunteer project to fix them within strict disclosure deadlines without providing patches or funding.

Future Outlook and Recommendations

Stenberg expects AI to continue augmenting every aspect of development. He urged projects to experiment with defenses against junk reports and AI scrapers, such as vetted-reporter "secret clubs" or stricter submission requirements — even if those measures conflict with traditional open source openness. Ultimately, he stressed that the problem is not AI itself but human choices: the same tools that enable "terrorist reporting" and bandwidth-hogging crawlers also significantly improved cURL's code security. Whether they benefit or harm depends on the judgment of maintainers, companies, and communities.