OpenAI announced higher security bug bounties on November 2, 2026, increasing rewards for independent researchers who report eligible software flaws. Researchers with pending reports or qualifying recent awards may receive more for work already submitted.
A bug bounty rewards eligible security findings reported to a company. The revised terms are live, effective November 2, according to OpenAI’s Bugcrowd notice.
Which existing reports qualify?
Still unawarded: Previously submitted reports receive the new reward ranges if a bounty has not yet been awarded.
Recently awarded: Bounties awarded in the 30 days before the November 2 announcement qualify for a difference payment when the revised base reward and loyalty bonus produce a higher amount.
For already-awarded reports, the retroactive promise covers base rewards and loyalty—not quality or campaign bonuses. The announcement does not define an exact time-of-day cutoff.
That makes award status and award date the useful checks in an existing report record, not submission date alone. OpenAI says it will review recent awards and notify eligible researchers of an additional payment; its notice gives no payment deadline.
What the reward ceiling means
Core-product groups covering ChatGPT, Codex CLI and the API platform list these base ranges in US dollars: medium $1,500–$3,000; high $3,000–$7,000; critical $10,000–$30,000. Other target groups use different schedules.
Exceptional critical findings remain eligible for a base award of up to $100,000; that exceptional limit is not new.
Eligible bonuses multiply: up to 1.15× for report quality, 1.30× for a priority campaign and 1.15× for the highest loyalty tier. Loyalty requires two lifetime valid submissions, with tiers based on impact points over the preceding 12 months. The policy also lists a low-quality multiplier of up to 0.85×. Rewards remain discretionary.
Illustrative maximum: $100,000 × 1.15 × 1.30 × 1.15 = $171,925, or roughly $172,000. This assumes the exceptional base and all three maximum bonuses. It is not typical earnings or a disclosed payment: OpenAI labels its example scenarios and historical awards as illustrative.
Which flaws count?
The Security program excludes model-only jailbreaks, hallucinations and prompt-based safety bypasses. A model-assisted report needs a separate failure of an enforced control in an eligible OpenAI product or OpenAI-controlled integration. Persuading an agent to misuse access it already has is not enough.
Some abuse risks instead fall under OpenAI’s separate Safety Bug Bounty, which has its own eligibility rules. General content-policy bypasses are not automatically eligible there either; the new Security reward schedule should not be assumed to apply.
Before starting new research, check the exact product’s live scope and reward group. Use authorized accounts and controlled data; higher rewards do not grant permission to test third-party systems.