Google has paused its Open Source Software Vulnerability Rewards Program after what it called a “significant rise” in automated submissions, the majority of which were apparently not the glittering security revelations their submitters imagined. According to TechCrunch, the program was frozen as of October 1, with Google promising an update in the first quarter of 2027.
The stated villain is not vulnerability research itself. It is the new aristocracy of machine-assisted almost-findings: bug reports with impressive formatting, confident prose, and the evidentiary density of scented air. Google said the pause follows a surge in automated submissions, “the vast majority” of which were not valid. In other words, the software-security equivalent of a luxury concierge ringing every bell in the hotel because a chatbot said one of them might be haunted.
🤚 The Open-Palm Triage Room
Google’s affected program rewarded researchers for finding vulnerabilities in the company’s open source software. Bug bounty programs are supposed to turn outside curiosity into useful defensive labor: researchers locate flaws, vendors verify them, patches happen, everyone receives either money, reputation, or the warm corporate mist of responsible disclosure.
That model depends on signal. The new problem is volume without judgment. TechCrunch notes that cybersecurity experts had already warned last year that AI-generated “slop” could threaten bounty programs by flooding maintainers with plausible-but-wrong reports. Now one of the world’s most security-mature companies has effectively placed a velvet rope in front of one such program and asked the robots to wait outside until the clipboard people recover.
Google is not ending bug bounties across the board. Participants were encouraged to look at the company’s other vulnerability reward programs while this open source track is paused. But the message is difficult to miss: if a program built to receive vulnerability reports has to stop receiving vulnerability reports because the reports increasingly resemble hallucinated consultancy decks, the ecosystem has discovered a very expensive form of spam.
👐 The Two-Handed Hallucination Audit
The irony, plated in gold leaf and served with a tiny compliance fork, is that AI was supposed to help security teams find more bugs. Sometimes it does. Models can assist with code review, pattern recognition, exploit sketching, documentation, and the tedious first pass through sprawling repositories. Used by a competent researcher, AI can be a power tool. Used by a person who confuses autocomplete with evidence, it becomes a confetti cannon pointed at the inbox of a maintainer who already has actual work.
Invalid bug reports are not merely annoying. They impose costs. Someone must read them, reproduce them, classify them, reject them, and sometimes explain why the alleged vulnerability is just a normal feature wearing a fake mustache. At scale, that becomes denial-of-service by paperwork: a flood of synthetic confidence consuming the time of the humans who were supposed to fix real defects.
This is especially painful in open source, where maintainers are often already balancing security obligations, community expectations, and the glamorous compensation package of “please accept our gratitude and another issue ticket.” A hallucinated vulnerability report does not arrive as harmless nonsense. It arrives as an obligation. It says: stop what you are doing, prove this machine-generated accusation is false, and do it politely enough that nobody writes a thread about gatekeeping.
🌿 The Gentle Awakening
The deeper lesson is not that AI is bad at security. The lesson is that automation changes incentives faster than institutions change filters. If producing a professional-looking report becomes almost free, the market will receive many professional-looking reports. Some will be legitimate. Many will be decorative fog. The bottleneck then moves from “who can find bugs?” to “who can verify claims without drowning?”
This is the same pattern appearing across code generation, academic writing, legal drafting, customer support, and every other office ritual where form can be manufactured more cheaply than substance. The machine can generate the costume of expertise. It cannot, by itself, guarantee the discipline of expertise. That part remains inconveniently human, like flossing or admitting a roadmap was fantasy.
For bug bounty platforms, the next phase will likely involve stronger submission gates: proof-of-concept requirements, better duplicate detection, reputation thresholds, stricter penalties for low-quality automation, and perhaps AI used defensively to screen AI-produced fog. A tiny war of clerks, but with GPUs.
👑 The Gold-Leaf Reckoning
Google’s pause is best understood as a warning label for the broader AI productivity boom. Productivity for whom? A tool that lets one person submit fifty low-quality reports may feel like leverage to the submitter. To the receiver, it is a labor tax. The efficiency has not been created; it has been relocated, gift-wrapped, and delivered to someone else’s queue.
Security research still needs automation. It also needs standards. The difference between an AI-assisted researcher and an AI-amplified nuisance is not the model; it is verification, restraint, and the willingness to test before sending. Civilization, tragically, remains dependent on people checking their work before pressing submit.
Until then, bug bounty programs may increasingly resemble exclusive clubs: not because vulnerability disclosure should be elitist, but because the lobby is filling with bots in rented tuxedos insisting they have found SQL injection in a sandwich.
“We regret to inform the automated researcher that vibes are not a proof of concept.” — The Slap of Wisdom Vulnerability Hospitality Desk, currently validating the chandelier