OpenAI boosts bug bounties 4x for its next security sprint

When you are building the most scrutinized software infrastructure on the planet, regular code reviews and standard annual penetration tests are no longer enough.

OpenAI announced plans for its next internal security sprint, teaming up with security firm Cathedral to stress-test its systems. Researchers from Cathedral will work alongside OpenAI internal teams, deploying the company's own frontier models to probe defenses, challenge internal assumptions, and track down hidden weaknesses across its platforms.

At the same time, OpenAI is putting serious cash on the table for independent researchers. It is overhauling its Security Bug Bounty program and quadrupling the payouts for critical vulnerability findings to more than four times previous levels, making it clear that finding critical flaws before external adversaries do is top priority.

Putting 25% of engineers on bug-hunting duty

This aggressive push is built directly on playbooks tested in earlier internal drills.

During its last sprint, OpenAI dedicated 25% of its entire product engineering roster solely to finding and fixing system vulnerabilities. Those engineers relied on Codex to investigate suspicious findings, reproduce live bugs in isolated environments, and prep patches for internal review.

OpenAI's underlying rationale is blunt: traditional cyber defenses simply cannot keep up anymore. Recent high-profile attacks across the tech sector show how quickly automated and AI-enhanced threats materialize. If defenders want to keep their defensive window open, they have to run continuous, automated loops that flag and fix problems around the clock.

A $1B subsidy to arm frontline defenders

Beyond patching its own backyard, OpenAI is attempting to institutionalize this approach across the broader ecosystem.

Through its Daybreak initiative, the company committed $1 billion in subsidized access over six months for frontline cyber defenders who protect essential digital services and critical infrastructure. It is also giving those teams access to its most capable models.

To help organizations replicate the setup, OpenAI is publishing its architecture, operational lessons from the last internal sprint, and a blueprint designed for industry and government defenders.

Why it matters

We think this is a pragmatic pivot for OpenAI, moving the conversation away from vague philosophical safety debates toward down-and-dirty infrastructure hardening.

For independent bug hunters, quadrupling top payouts provides an immediate financial incentive to pound on OpenAI's perimeter. For underfunded security teams guarding critical public infrastructure, a slice of that $1 billion in subsidized model compute could make AI-assisted vulnerability patching genuinely accessible for the first time.

For the cybersecurity industry at large, OpenAI's move sets a fresh standard: auditing code using frontier models is transitioning from an experimental novelty into an operational baseline.

Yet the approach carries its own risks. Handing over significant portions of vulnerability reproduction and patch generation to AI models means any hallucinations or blind spots could accidentally introduce stealthier vulnerabilities into production. It looks like the arms race between AI-assisted attackers and AI-assisted defenders has officially started, and human engineers will have to stay sharp to verify every single fix.