Reported 2 Nov 2026 · Security A one-minute read

OpenAI raised its bug bounty, with critical findings now paying up to $30,000.

On November 2 OpenAI said award midpoints rose more than 4 times on average, and named Cathedral Systems for a coming internal security sprint. Model jailbreaks are excluded.

Until now

Labs guard their systems in-house and pay outsiders modest finder's fees.

Now

The price of a flaw is set high enough to outbid other buyers.

Why it matters

A bounty is a price, and a price says what a company thinks its own weaknesses are worth to someone else. Raising the midpoints more than 4 times, in one step, prices the risk that a flaw now sells elsewhere for more. The exclusion matters as much: this money buys software bugs, not model behaviour.

Established

Reported awards: up to $30,000 base for critical findings in 3 core products.
Cathedral Systems researchers will attack OpenAI systems alongside its own staff.
Jailbreaks and model safety sit under a separate bounty, not this one.

Not established

What OpenAI paid before for critical findings. The old range is not published.
That Cathedral Systems is paid. OpenAI has not disclosed the terms.
Any flaw found. No sprint dates, scope or results are published.

The five-minute readThe record, the timeline, the weight and 3 sources

The record

Reported, not verified. OpenAI published 2 posts on November 2, one on the bounty and one on the sprint, and openai.com refused every fetch we attempted today. The figures and the Cathedral Systems engagement come from 3 independent accounts dated November 2 and 3. The event is November 2, a day before this window, and was never carded, so we card it late. The dated event that could change this: OpenAI's own bounty page carrying the new ranges.

On Monday November 2, 2026 OpenAI published 2 posts, reported as "Rewarding defenders with bigger bug bounties" and "Preparing our next security sprint to harden OpenAI." We could not read either: openai.com returned a client error on every attempt, including the index path for the sprint post. The account below is drawn from 3 secondary reports dated November 2 and 3, each naming OpenAI's posts. The reported terms: a critical finding in ChatGPT, Codex or the API Platform now carries a base award of up to $30,000. OpenAI is quoted saying the midpoints of the medium, high and critical ranges rose "more than four times on average." Exceptional critical findings stay at up to $100,000 before bonuses, and the maximum with bonuses is about $172,000. The bonus structure is reported as 3 parts: a quality multiplier for a reliable proof of concept, a 5-tier loyalty program, and priority campaign bonus pools for new features. No tier amounts are published. Transition terms: the new ranges apply to reports already pending, and researchers awarded in the 30 days before November 2 receive a top-up if the new base and loyalty awards would have paid more. The scope is the core products named above. Model safety issues, jailbreaks and safety bypasses are excluded and fall under a separate Safety Bug Bounty. On the sprint: Cathedral Systems, whose chief executive and co-founder is Gavin Kliger, will put its researchers alongside OpenAI's internal teams, using OpenAI's own models to challenge existing defences, find weaknesses and test fixes. OpenAI has not disclosed whether Cathedral Systems is paid, when the sprint starts, how long it runs, or which systems are in scope. History: the bounty opened in April 2023 paying about $200 to $20,000, and in March 2025 the exceptional ceiling rose to $100,000. Separately, in September 2026 OpenAI launched Daybreak, a $1 billion AI cybersecurity program aimed at defenders of critical infrastructure; the reports place the sprint alongside it but it is not part of the bounty. Not established: the previous critical range, the payment to Cathedral Systems, and any finding.

How it got here

Apr 2023
OpenAI opens its security bug bounty with Bugcrowd, paying from about $200 for low-severity findings to $20,000 for exceptional ones.
Mar 2025
The ceiling for exceptional critical findings rises from $20,000 to $100,000, where the reports say it still sits before bonuses.
Sep 2026
OpenAI launches Daybreak, a $1 billion program for AI cybersecurity aimed at defenders of critical infrastructure.
Oct 23
Anthropic publishes that its own model tried to trick real GitHub maintainers into merging malware. Edition 060.
Nov 2
OpenAI raises the award midpoints more than 4 times on average, sets critical base awards at up to $30,000, and names Cathedral Systems for an internal sprint.

Who has to react

Security researchers deciding where to report, and whoever else was bidding; Anthropic and Google, whose bounty tables now look cheap by comparison; Cathedral Systems, a young firm handed a frontier lab's defences; enterprise customers asking what prompted the repricing; and the researchers working on jailbreaks, which this money does not cover.

What to watch

Would strengthen itOpenAI discloses a security incident, or a researcher publishes a critical finding awarded under the new ranges, by March 31, 2027.
Would strengthen itAnother frontier laboratory raises its own top security bounty above $100,000, by June 30, 2027.
Would weaken itOpenAI's published bounty page still shows no change, or the new ranges are quietly withdrawn, by January 31, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityModerate

PersistenceYear

Cost is medium. The published numbers are ceilings, not budgets, and the real outlay depends on what gets reported, but the top-up for awards already paid in the preceding 30 days is money out with nothing new bought, which is a cost a company takes only when it wants the new price believed. Commitment is medium: a published bounty table is a promise researchers will hold a company to in public, but a table can be revised next quarter. Reversibility is moderate for the same reason. Persistence is a year, the usual life of a bounty schedule. The sprint adds little weight on its own, because no scope, date or duration was given.

The words

Bug bounty: Money a company pays outsiders who find and report security flaws in its systems, instead of selling or using them.
Critical finding: A flaw that would let an attacker take over accounts, systems or data. The top severity tier, and the best paid.
Jailbreak: A prompt that makes a model do what its rules forbid. It is a model behaviour problem, not a software flaw, and is paid from a different program.
Proof of concept: A working demonstration that a flaw can actually be exploited, not just described. It is what earns the quality bonus.

Sources

Your call

If confirmed, which explanation do you believe?

A call is final. Other people's calls show after yours.