AI Has Made Bug Bounties Cheaper to Flood Than to Fix

We tend to equate automation with progress: if AI can produce more, the system must become more productive. Yet Google’s decision to pause its Open Source Software Vulnerability Rewards Program reveals a less convenient truth. When generation becomes almost free, verification becomes the scarce capability.

Google paused the program effective October 1 after automated submissions surged. The company says most were invalid, with engineers and maintainers overwhelmed by hallucinated or unusable reports; it expects to provide an update in the first quarter of 2027.

The Signal Is an Architecture Problem

This is not merely a moderation problem. It is a classic systems-engineering failure: the intake pipeline was designed for a world where a submission was expensive enough to be relatively selective. Generative AI broke that economic assumption. A machine can now manufacture syntactically convincing vulnerability claims at a scale that human adjudication cannot absorb.

The deeper principle is simple: AI scales production, not truth. Plausibility is no longer evidence.

In many enterprises, the same pattern will appear in compliance filings, incident tickets, research annotations, partner integrations, and knowledge-base updates. If downstream teams treat every machine-generated artifact as meaningful input, they risk creating an overloaded control plane-and losing sensitivity precisely when it matters most.

Verification Is the New Competitive Layer

The response cannot simply be “add more AI reviewers.” LLMs can cluster duplicates, extract reproduction steps, compare advisories, and flag missing evidence. But probabilistic systems must not quietly become the authority deciding whether a real vulnerability is accepted. They should accelerate triage, while independent experts retain final judgment.

Architecturally, every high-volume submission channel needs an evidence pipeline: authenticated identity and reputation; malware-safe sandboxing; deterministic reproduction; version-aware validation; duplicate detection; risk-based queues; and an auditable decision trail.

Provenance matters too. Was the report generated, translated, or modified by a model? An AI-disclosure label may help, but unreliable AI-detection scores should never be treated as a trust control. Reproducibility is the stronger signal.

“Human in the Loop” Must Mean Something

The phrase “human in the loop” has become decorative in many AI systems. A human who receives hundreds of unsupported claims, lacks time to reproduce them, and can only approve or reject a model’s recommendation is not really in control.

The scarce roles are now bug hunters, security engineers, and adjudicators capable of separating a novel exploit from an invented one. This is not a case for replacing experts with agents. It is for using agents to remove low-value inspection work while giving experts better evidence, clearer priorities, and explicit authority to reject weak submissions.

The Economics Must Change as Well

A bounty program is also a market. If reward decisions become slower because low-quality volume consumes reviewer capacity, the program suffers adverse selection: legitimate researchers wait, good reports age, and participation declines. Queueing theory is now part of cybersecurity strategy.

Google’s pause is therefore rational, but pausing alone is not the long-term design. Programs need quality-weighted triage, response-time commitments by severity, useful rejection feedback, reviewer-capacity planning, and abuse controls that distinguish coordinated spam from genuine but inexperienced reports.

The right metric is not submissions generated; it is validated risk delivered per hour of expert attention.

A Practical Agenda for CTOs

Leaders should treat AI-generated reports, tickets, documents, and code as untrusted external input-not as completed work. They must define machine-verifiable evidence requirements before automating intake. Furthermore, they should use models for summarization, clustering, and test preparation-not final truth. It is essential to track validity rate, reviewer cost, duplicate load, time to disposition, and escaped-risk recall. Organizations should also publish closed-loop feedback so contributors learn what constitutes reproducible evidence, and they must fund verification capacity as seriously as generation tooling.

For students and young researchers, this is a useful career signal. Coding help, frameworks, and cybersecurity-as-a-service are not durable differentiators. Systems thinking, threat modelling, experimental discipline, and the ability to prove a claim will be.

The next advantage will not belong to those producing the most AI output. It will belong to those building systems that can cheaply determine which output deserves to be believed.

---

About the Author: Sanjeev Sarma is the Founder Director and Chief Software Architect at Webx Technologies. With a core focus on Generative AI integration, Cloud-Native Scalability, and Enterprise Software Architecture, he has spent over two decades driving digital transformation across Northeast India and beyond. Beyond his corporate leadership, Sanjeev is deeply invested in shaping the future of the IT industry. He serves as an Industry Expert on the Board of Studies for Assam Don Bosco University’s School of Technology, advises state technology committees, and actively mentors emerging tech startups at STPI. He brings a unique, dual perspective of high-level enterprise execution and future-ready academic curriculum development.