Rewarding defenders with bigger bug bounties

OpenAI is raising its Security Bug Bounty rewards sharply. The midpoint of reward ranges for medium, high and critical findings is more than quadrupled on average. A critical bug in ChatGPT, Codex or the API Platform can now earn up to $30,000, and exceptional critical findings can reach up to $172,000 with bonuses.

The company is adding multipliers for high-quality reports, a five-tier loyalty program, and bonus pools for newly launched features. Unpaid reports already submitted get the new rates, and recent awards from the past 30 days get topped up.

As AI products hold more of people's work and conversations, the price of a missed bug goes up. Paying more to find them first is a reasonable response.

We're in favor of this. Bug bounties work best when the money matches the effort, and the old figures OpenAI shows in its own examples, like $6,500 for a one-click account takeover, were low for a product used at ChatGPT's scale. The new example for that same bug comes to more than $51,000 with bonuses.

The quality multiplier is the most thoughtful piece. A clean proof of concept that isolates the bug and shows its impact saves the security team real time. Paying for that discourages the flood of thin, speculative reports that clog many programs, a problem that's likely to grow as more people point AI tools at targets.

The retroactive payments are a good-faith move, and they matter. Researchers remember how a program treats them.

A few cautions. The headline numbers include stacked bonuses, and the company says its examples are illustrative, so most payouts will be lower than the top figures. Loyalty tiers can also nudge researchers to stick with one vendor. That's good for OpenAI, but the wider ecosystem benefits when skilled people spread their attention around.

Our read: a well-designed upgrade. What matters next is whether triage speed and fairness keep up with the bigger checks.

OpenAI positions the bounty as complementing its Defense Factory work, where engineers use its models to find, reproduce and fix weaknesses. It was announced alongside a new internal security sprint with outside firm Cathedral. The security team will also be at PROMPT2PWN in Oakland on November 20–21.