OpenAI quadruples bug bounties, adding a roughly $172,000 ceiling

By James Calloway

OpenAI will pay external security researchers up to roughly $172,000 for vulnerability findings after quadrupling its bug bounty payouts, the cash rewards offered for reporting software flaws. The company says critical flaws in core products, including ChatGPT and Codex, can earn up to $30,000, while exceptional findings can reach $100,000.

Quality, loyalty, and campaign bonuses can raise the total to roughly $172,000. That headline is a maximum rather than the standard award, and the announcement does not explain how the bonuses combine.

What changed in OpenAI's payout schedule?

OpenAI announced the fourfold increase as part of a broader effort to strengthen security around its products. The revised schedule sets a $30,000 ceiling for critical vulnerabilities in core products such as ChatGPT and Codex and a $100,000 ceiling for exceptional findings, then adds several possible bonuses.

The announced figures establish ceilings, not guaranteed payments. A report must meet the relevant severity standard, although the source excerpt does not define those standards. Nor does it explain whether the $30,000 and $100,000 limits include or precede the named bonuses.

That accounting distinction carries weight. If “quadrupled” refers to a base award while the $172,000 figure includes add-ons, the two headline numbers describe different calculations rather than competing maximums. The source does not resolve that point.

Takeaway: OpenAI has made the reward ceiling much larger, but the announcement leaves the calculation largely unstated.

How does the program compare with Google, Apple, and Microsoft?

The Decoder places OpenAI’s new program in the same ballpark as Google’s general vulnerability rewards program. It also says OpenAI remains below the top-tier payouts Apple and Microsoft offer for operating-system-level exploits.

The comparison supports a broad claim about competitiveness. It does not show equivalent coverage, because the source provides no payment ranges or eligibility rules for Google, Apple, or Microsoft. Operating-system exploits and vulnerabilities in ChatGPT or Codex also involve different technical and commercial contexts.

The missing numbers make direct price comparisons impossible. A reader can conclude that OpenAI intends to compete for security research, not that its program matches any rival on every product class or severity level.

Takeaway: The headline is competitive, but the source supports only a rough market comparison.

What do the new loyalty and campaign bonuses do?

OpenAI added a five-tier loyalty program and time-limited campaigns with boosted payouts tied to new feature launches. The first appears intended to reward researchers who return to the program, while the second concentrates attention on issues associated with new releases.

The source does not name the five tiers, state their thresholds, or explain how long researcher status lasts. It also gives no dates, products, or campaign terms for the feature-linked windows. Those details determine whether a higher advertised rate is realistically available to an individual report.

The company also made the increase retroactive for a defined group. Anyone who reported a vulnerability in the previous 30 days will receive the difference between the old and new rates. The stated policy could prevent researchers from delaying reports while waiting for richer terms.

Takeaway: Bonuses broaden the incentive structure, but researchers cannot calculate expected rewards without the missing tier and campaign rules.

What is the new security sprint?

OpenAI said it is preparing a “security sprint” in which the security firm Cathedral will work with the company’s engineers to stress-test its systems. The exercise places an outside security firm inside the same effort as the product team, rather than relying only on the public bounty channel.

During the previous sprint, 25 percent of OpenAI’s product engineers worked on “finding vulnerabilities,” according to the company. In practical terms, that means one in every four product engineers joined the internal search. The figure comes from OpenAI, and the source gives no employee count, sprint duration, target list, or outcome data.

The source offers no independent audit of the participation figure. The announcement also does not say whether the sprint concentrates on ChatGPT, Codex, both products, or other systems. Its immediate purpose is narrower: engineers and a security firm will stress-test the systems, while outside researchers retain a separate incentive to report findings.

Takeaway: The internal sprint adds a second line of defense, but its scale and results remain mostly unquantified.

Why do the new incentives matter?

The program uses three levers at once: a higher payout ceiling, an exceptional-finding award, and extra rewards for quality, loyalty, and campaign participation. For outside researchers, the top possible payout is roughly $172,000. That figure sits $72,000 above the $100,000 exceptional award, but the source does not disclose the bonus amounts needed to close the gap.

The structure also gives OpenAI a way to shape research attention. Feature-linked campaigns can direct testing toward newly released functionality, while the broader bounty remains open to discoveries elsewhere. The five-tier program appears designed to retain researchers over time, though the source does not confirm that repeat participation is the sole criterion for advancement.

The stakes are practical for both sides. The likely result is a broader pool of potential testers around product changes. Researchers gain more reasons to examine those changes closely, but a maximum payout says little about frequency, severity distribution, or the likely median award.

The program’s true market value will depend on more than the ceiling. It will depend on whether rewards arrive quickly, whether the criteria remain clear, and whether exceptional findings are genuinely exceptional. None of those operating details appear in the source.

Takeaway: OpenAI is buying broader scrutiny and more targeted release testing, not promising routine $172,000 checks.

What has not been disclosed?

The source leaves several questions that determine how researchers should interpret the announcement:

• It does not define “critical” or “exceptional,” including which ChatGPT or Codex behaviors qualify.
• It does not provide the old payout schedule or a formula showing how the fourfold increase was calculated.
• It does not explain the bonus amounts, whether they can be combined, or what produces the roughly $172,000 total.
• It gives no five-tier loyalty thresholds, campaign dates, launch targets, campaign duration, or effective date for the revised rates.
• It omits payment timing, reporting requirements, disclosure rules, and any process for correcting a reported vulnerability.
• It supplies no Cathedral engagement scope, sprint schedule, discovery count, remediation record, or independent review of the 25 percent participation figure.

That list limits how confidently anyone can compare the program with Google’s general rewards or Apple’s and Microsoft’s operating-system payouts. The source supports the existence of larger maximum awards and a new testing sprint, not a full account of their practical value.

Takeaway: The announcement establishes ambition and headline numbers, but not the terms needed to audit performance.

What should readers watch next?

The first test is whether OpenAI publishes enough detail to make the fourfold increase auditable. Clear severity definitions and bonus formulas would let researchers compare expected awards, while vague labels would leave the program’s biggest figures largely rhetorical.

The second test is whether feature campaigns name the launches they target and remain active long enough to test meaningful behavior. A boosted rate attached to an unspecified window offers little planning value. The five-tier system also needs transparent advancement criteria if it is meant to create durable participation.

The third test concerns internal testing. The company has quantified participation through the 25 percent figure, but it has not explained what the previous sprint found or changed. Reporting outcomes, even in aggregate, would show whether an engineer-led sprint complements the external bounty rather than duplicating it.

For now, OpenAI has set a large ceiling and broadened the financial incentive for outside vulnerability research. The missing program mechanics will determine whether that ceiling changes researcher behavior or remains mostly marketing.

Takeaway: The next useful evidence is measurable: published rules, named campaigns, and disclosed sprint results.