Epizódjegyzetek
For three decades years, software makers have paid bug bounty hunters to break their code so they can make it more secure. Recently, some software companies have decided that bug bounties programmes aren't useful any more: Curl shut its bounty in January, and Intel suspended its own in September. And the discovery of the PHP flaw that sent websites offline in October came from researchers working with AI, not from a bounty hunter's report.
Tobias Kretschmer (Imperial College London, CEPR) and his co-authors followed 86 programmes on Bugcrowd to discover how they work best. Do you run closed programmes, open only to invited hackers, or open them to all comers? And what happens now that AI is making a bug report cheaper to produce? Tim Phillips asks Kretschmer how well bug bounties are working.
The research behind this episode:
Zrahia, Aviram, Sarit Markovich, and Tobias Kretschmer. 2026. "Platform Openness: Evidence from Bug Bounty Programs." CEPR Discussion Paper 21970 (gated).
To cite this episode:
Phillips, Tim, and Tobias Kretschmer. 2026. "How well do bug bounties work?" VoxTalks Economics (podcast).
About the guest
Tobias Kretschmer is Professor of Strategy, Technology and Organisation at Imperial College Business School and a Research Fellow at the Centre for Economic Policy Research. He joined Imperial in 2025 from LMU Munich. His research spans competitive strategy in digital industries, platform markets, and the digital transformation of organisations.
Research cited in this episode
Bug bounties. Netscape launched the first technology bug bounty on 10 October 1995, offering rewards for flaws found in the Navigator 2.0 beta. Jarrett Ridlinghafer, a technical support engineer, proposed the idea after he saw enthusiasts finding bugs for free. In this paper's data, the average reward for a unique vulnerability is about $566 in private programmes and $209 in public ones. High-priority flaws average about $1,440.
Bugcrowd connects organisations with independent security researchers. It gave the authors anonymised records of more than 480,000 submissions to 2,483 programmes, from mid-2012 to mid-2021. The data predates the widespread use of AI-generated submissions, so the paper measures the effect of openness, not the effect of AI. The authors argue that AI sharpens the same problem.
Triage. Each submission is classified as rejected, duplicate, or unique. A unique report identifies a flaw nobody has reported before. Only the first valid report generally earns a reward, so a duplicate can be good work that arrived late. It must still be reviewed.
Wisdom of the crowd. Francis Galton's "Vox Populi" (Nature, 1907) reported that the median of about 800 guesses at an ox's dressed weight, made at a Plymouth livestock show, was remarkably close to the true figure. Eric Raymond gave software its version in "The Cathedral and the Bazaar": "given enough eyeballs, all bugs are shallow". Maillart and co-authors tested that idea on bug bounties in…
Weboldal: audioboom.com

