OpenAI paid $6,500 for a three-bug chain that reached its private monorepo

Hacktron's team found Claude Opus 4.8 unreliable for the exploit work ^[16]. Opus 5, released on July 24, produced a working ARM64 exploit within hours, and the team adapted it for x86-64 and jemalloc environments ^[17]. Memory-corruption exploit development is specialist work that has historically taken skilled humans weeks ^[8]. The published account does not say how much of the chain the model drove, how much the team did, or how large the team was. Before reading the Opus 5 result as the end of a talent constraint, note who was at the keyboard: researchers affiliated with the security startup Hacktron AI ^[6].

OpenAI paid $6,500 for a report that ended with a pull request inside openai/openai ^[1][4]. Spread across the under-72-hour window, that is about $2,170 per calendar day ^[14][7]. For a program owner the constraint sits elsewhere. Coinbase has warned that bug reports could triple in the near future as AI floods disclosure programs with noise ^[12]. A program pays per valid report and pays humans to read the invalid ones. The payout schedule sets only the first of those two prices.

The volume evidence sits alongside that single chain. Chainalysis reported this week that attackers are posting malware instructions to public blockchains 440% more often than a year ago, with daily malicious onchain writes climbing from 2.06 to 11.1 ^[9], a ratio of roughly 5.4 to one ^[15]. Those figures count malware staging on public ledgers, a different thing from intrusions at frontier labs. Chainalysis tied the jump to mid-2025, when open-weight Chinese models launched without meaningful guardrails against writing malicious code, and called the technique blockchain dead drops ^[10]. By the second quarter of 2026, state-linked operators from North Korea and Iran accounted for roughly two-thirds of new activity and about half the total ^[11]. Researchers tracking North Korea's Kimsuky found Ollama, GPT4All and Msty installed on the group's infrastructure alongside AI-generated phishing decoys ^[13].

The timeline explains part of why the price looks low. OpenAI fixed the vulnerability the same day it was reported, and Discourse published its advisory on July 28 with a CVSS severity of 8.8 ^[5]. The episode became public on Sept. 17, when the Wall Street Journal reported it ^[18], 54 days after the report landed ^[19]. OpenAI set the payout while it and Hacktron were the only parties who could see the chain.

The decision in front of a program owner this quarter is narrow: a rate card that separates a chained path to source access from a single finding, and enough triage capacity for the volume Coinbase describes ^[12]. The longer question is whether a payout sized for scarce human exploit work still clears a market where a three-stage chain into a frontier lab takes under three days ^[7]. Opus 4.8 could not do that task, and Opus 5, one release later, could ^[16][17].