Bug hunting can now earn $172,000 at OpenAI

Faced with AI-boosted cyberthreats, OpenAI is preparing an audit with Cathedral Systems and raising its bug bounties to researchers to as much as $172,000.

A critical vulnerability discovered in ChatGPT, Codex, or the OpenAI API can now earn its author up to $172,000, bonuses included. This exceptional increase in rewards accompanies the preparation of a new security testing campaign led with external researchers. The company intends to use its own models to identify weaknesses in its infrastructure before they can be exploited by attackers.

The new security operation will bring together internal teams and Cathedral Systems, a company specializing in vulnerability research. Its researchers will have access to OpenAI’s models to examine the platform’s defenses, reproduce the flaws found, and help fix them. The schedule and duration of this campaign, which is still being prepared, have not been specified. Payouts multiplied for critical flaws

The overhaul of the bug bounty program applies to independent researchers who responsibly report vulnerabilities in the company’s products. The median reward amounts for medium, high, and critical severity flaws are more than quadrupled on average.

For the main products, including ChatGPT, Codex, and the API platform, a critical vulnerability can now be rewarded with up to $30,000. Exceptional findings remain eligible for a base bounty of up to $100,000, which can be raised to about $172,000 thanks to various bonuses. This ceiling therefore does not represent the usual reward for a critical flaw.

Three new mechanisms are being added to the program. Report quality will be given greater weight, especially when reports include a reproducible proof of concept and a precise description of the impact. A five-tier loyalty program will reward regular researchers, while priority campaigns will offer additional multipliers for certain recently introduced features.

The new terms will also apply to reports already submitted but not yet rewarded. For bounties paid in the thirty days preceding the announcement, an additional amount may be granted if the new scale proves more favorable. When Codex takes part in penetration testing

This campaign extends a previous internal operation during which 25% of OpenAI’s product engineers were mobilized to search for and fix vulnerabilities. The teams used Codex to review reports, reproduce issues in isolated environments, and prepare fixes that were then submitted for human review.

The approach consists of integrating artificial intelligence into the entire defense process, from discovering a weakness to fixing it. A model capable of exploring a codebase, understanding interactions between multiple components, and testing different hypotheses can accelerate certain investigations that are usually lengthy. The same capabilities can, however, also be used to search for flaws for malicious purposes, which reinforces the importance of access controls and terms of use.

The use of Cathedral Systems introduces outside expertise into this setup. Unlike a fully internal review, this collaboration is meant to confront the systems with other investigative methods and with hypotheses that the usual teams would not necessarily have considered. No results from this new campaign have yet been announced. A billion dollars in subsidized access for defenders

The initiative is part of a broader cybersecurity strategy built around Daybreak, an offering intended for professionals authorized to use advanced models for defensive operations. The company is seeking to extend these capabilities to organizations responsible for essential services, including water and power networks, local governments, and financial infrastructure.

As part of Daybreak for Frontline Defenders, a $1 billion commitment has been announced to subsidize access to models, training, technical support, and partnerships. This allocation, intended to be used over six months, does not correspond to a direct $1 billion investment in the internal testing campaign alone.

The company also defends the concept of a defender’s window, a period during which cybersecurity actors could use advances in AI to strengthen their systems before those capabilities spread more widely among attackers. This hypothesis rests on a temporary advantage for defenders, which cannot, however, be guaranteed by the mere availability of more capable models.

The new campaign with Cathedral Systems still has to be conducted, while the revised reward scale has already been published. The results of the previous internal exercise are not accompanied, in this announcement, by a numerical assessment of the vulnerabilities discovered and fixed. The next data points will make it possible to better assess what these methods actually contribute to the security of the infrastructure involved.