OpenAI Hikes Bug Bounty Payouts to $172,000 and Partners With Cathedral for AI Security Sprint

OpenAI is framing its latest security push as a necessary evolution in digital defense, but the underlying reality is more pragmatic: as artificial intelligence models become highly capable of writing code, they simultaneously become adept at breaking it. To protect its own infrastructure from autonomous agent exploits, the company is dramatically increasing its bug bounty payouts.[1][7]

The company announced on November 2 that it is more than quadrupling the midpoint of its reward ranges for medium, high, and critical software vulnerabilities. A critical flaw discovered in core assets like ChatGPT, the Codex model, or the API Platform now commands a base bounty of up to $30,000.[1][2]

For exceptional critical findings—such as a significant compromise of a security boundary that allows an attacker to view arbitrary user messages—independent researchers can earn a base award of $100,000. With a newly introduced system of multipliers, that total payout can reach a maximum of $172,000.[1][3]

The payout structure provides concrete examples of how the multipliers stack. A medium-severity vulnerability, such as a read-only member uploading files to a project without write permission, previously earned $750. Under the new system, with a quality bonus and a Bronze loyalty tier multiplier, that same finding pays $2,415.[1]

For a high-severity flaw, like an insecure direct object reference that allows reading a file uploaded to another user's project, the base reward is $2,500. When combined with a quality bonus and a Gold loyalty tier multiplier, the final payout reaches $3,162.[1]

The highest tier of the loyalty program, dubbed the 31337 tier, offers the largest scaling bonuses for elite researchers. A one-click account takeover in ChatGPT, previously valued at $6,500, now triggers a $30,000 base reward that scales past $51,000 when priority campaign and top-tier loyalty multipliers are applied.[1]

AI-generated

Incentivizing Proof Over Speculation

The revised bonus structure deliberately ties richer financial rewards to reports that prove real-world exploitability rather than theoretical speculation. A new quality multiplier rewards a reliable proof of concept that successfully isolates the bug, shows exactly when it can be exploited, and demonstrates its material impact.[1][6]

OpenAI is also introducing a five-tier loyalty program that offers scaling bonuses for security researchers who consistently partner with the company. Additionally, short-term priority campaigns will feature dedicated bonus pools to encourage hackers to aggressively test newly launched features immediately after release.[1][3]

The payout hikes are effective immediately, and the company is making them retroactive to catch recent vulnerability submissions. "If you've already submitted a report and we haven't yet awarded a bounty, we'll apply the new reward ranges," OpenAI stated in its official release.[1]

For bounties awarded in the 30 days prior to the announcement, OpenAI will pay the difference if the finding would have earned more under the new base and loyalty structures. The move places OpenAI's payouts in the same tier as Google's vulnerability rewards, though still below the maximum bounties offered by Apple or Microsoft.[1][3]

The Cathedral Partnership

Alongside the financial incentives for independent hackers, OpenAI is preparing a new internal security sprint to actively hunt for vulnerabilities before external attackers find them. To execute the sprint, the company has partnered with Cathedral Systems, a cybersecurity startup founded by former government staffers.[4][7]

Cathedral Systems is a relatively new entrant in the cybersecurity sector, having been founded by former staffers of the Department of Government Efficiency. Despite its recent arrival, the firm has quickly established a reputation for fast and effective offensive security testing.[4]

Cathedral recently demonstrated its offensive capabilities by deploying artificial intelligence agents that successfully found security gaps in rival AI firm Anthropic's third-party business systems in under 12 hours. OpenAI is now bringing that same aggressive, model-assisted testing methodology inside its own walls.[4]

Cathedral researchers will use OpenAI's own models alongside internal engineering teams to probe the network for hidden weaknesses. The collaboration is designed to challenge existing protections and bring fresh, adversarial approaches to testing the company's platforms against automated threats.[5][7]

"OpenAI is setting an important example by partnering with outside security firms and using its own models to continuously test and strengthen its defenses," said Gavin Kliger, the chief executive officer and co-founder of Cathedral.[7]

Kliger noted that every technology company should be evaluating how to put these automated capabilities to work. He added that Cathedral is excited to combine its research with OpenAI's models to find weaknesses, test fixes, and support a posture of continuous, active defense.[7]

AI-generated

Automating the Defense Factory

The upcoming security sprint builds on a previous internal exercise that supplies the operating template for OpenAI's current security operations. During that earlier sprint, the company dedicated 25 percent of its product engineers entirely to finding and fixing software weaknesses.[6][7]

Engineers used the Codex model to investigate vulnerability reports, reproduce the flaws in isolated test environments, and prepare software patches for human review. According to OpenAI's architecture documents, the effort mobilized more than 250 people across over 100 distinct service areas.[5][6]

The automated approach yielded immediate operational results, closing 53 urgent or high-priority security issues on its first day alone. The system identified 37 percent of findings as duplicates and successfully reproduced 19.5 percent of the vulnerabilities at runtime.[6]

By automating the validation process, OpenAI recorded a remarkably low 0.81 percent false-positive rate after dynamic testing. Remediation was entirely Codex-based, resulting in a 0.53 percent rolled-back fix rate, indicating that the AI-generated patches were largely stable and effective.[6]

However, the sprint also exposed the limitations of relying purely on automated ticket management and AI-generated code. Follow-up checks revealed a critical gap between a patch being merged into the codebase and the fix actually being deployed across the entire production fleet.[5][6]

AI-generated

Closing the Defender's Window

OpenAI noted that ticket movement or accepted work does not inherently prove remediation. The verified-remediation stage requires concrete deployment evidence and independent checks after human review, proving that a merged patch does not guarantee a secure system until it is fully live.[6]

The company argues that traditional cyber defenses are no longer sufficient to protect against modern, automated threats. As artificial intelligence lowers the barrier to entry for attackers, organizations must build continuous defensive loops to find and fix weaknesses before malicious actors can exploit them.[7]

To extend these defensive capabilities beyond its own infrastructure, OpenAI is making its most capable models available to outside security teams through its Daybreak initiative. The program specifically targets resource-constrained defenders of essential services, such as water and electricity providers in the United States.[5][6]

OpenAI has committed $1 billion in subsidized access, training, and technical support over six months for Daybreak customers. As of September 2026, the company reported that thousands of defenders across 2,000 approved organizations were already utilizing the platform for authorized defensive work.[6][7]

The Daybreak initiative also includes a blueprint for empowering cyber defenders across both industry and government sectors. By sharing the architecture and operational lessons from its internal security sprints, OpenAI aims to establish a standardized framework for model-assisted vulnerability management.[7]

To further engage with the independent security community, OpenAI's security team will attend the PROMPT2PWN conference in Oakland on November 20 and 21. The event will allow researchers to test targets in person and discuss potential partnerships directly with the company's engineers.[1]

The planned security sprint and the bug bounty hikes represent two sides of the same overarching strategy. Outside researchers receive massive financial incentives to expose what internal testing missed, while the internal agent-first operation is built to rapidly validate and patch those exact weaknesses.[6][7]