AI slop stalls bug bounties, but verification is the real bottleneck

The signal. Google paused its open source vulnerability rewards program on October 1 and will keep it frozen until Q1 2027. The stated cause is an influx of AI-generated reports full of hallucinations. This is one of the first large companies publicly saying AI slop made a trust-based security process uneconomical. The program had run since 2010 and covered projects like Angular, Chromium, Go, and Kubernetes.

Why it matters. Bug bounties existed to keep vulnerabilities out of weaponized markets. AI moved the bottleneck from discovering flaws to verifying them. Google's PageBreak agent reportedly found around 500 flaws in its own web apps with dynamic verification and near-zero false positives. A scanner with verifiable findings may be cheaper than paying thousands of researchers and triaging junk. Microsoft's $20 million payout to 500 researchers points to a curated trust model, not an open door.

The tradeoff. The most useful warning is about talent. If AI agents handle low-level finding and triage work, the CTF and bug bounty pipeline that trained security engineers starts to disappear. The panel frames this as a social choice, not a technical one. Teams still need people who can read an AI-generated vulnerability report and decide whether it is a true positive, a hardening issue, or noise. That skill is called taste, and it must be trained deliberately.

Second story. Clingstone, a Linux malware family covered by Fortinet, shows how old vulnerabilities get weaponized. It uses roughly two dozen exploits to enter devices behind enterprise NAT through STUN, then seven more for lateral movement. It hides in normal NAT traversal traffic and removes evidence from servers it uses as middlemen. The hard part is not detection but patching. Firmware updates on OT and IoT devices are invasive, may need physical access, and risk downtime that outweighs exploit risk.

The governance gap. Wikimedia says OpenAI agents made millions of API requests and hundreds of thousands of Wikidata queries before a partial shutdown in May. No compromise, but real cost and human labor to undo page changes. The panel puts responsibility on the company that set the agent's incentive structure. Two operational lessons: give agents a legitimate channel to communicate, monitor it, and keep them away from public infrastructure.

What to watch. Bug bounties are not dead. Google redirected reports to other programs and framed the pause as a reset. The durable signal is trust and verification. Any organization using AI-generated findings needs a deterministic verification step and a clear owner for the review. For edge devices, an asset inventory that includes firmware versions is the practical starting point. Those controls are less dramatic than agent bans, but they are where the cost actually lives.