2026 OpenAI security bug bounty: 4x critical rewards vs Cathedral tests

Verify OpenAI's revamped Security Bug Bounty before committing.

Compare like-for-like totals and terms; verify any claimed reward multiplier against the live program page rather than the headline.

How It Works

At its core, the mechanism for this pairing is a controlled test loop: a partner is brought in to probe a system under an agreed scope, findings are submitted through a defined reporting path, and those findings are evaluated against the program's stated terms before any action is taken. According to Preparing our next security sprint to harden OpenAI (OpenAI via news.google.com), OpenAI has stated it is preparing a security sprint to harden OpenAI, which points to outside testing being used as one part of that effort. For verification, the check is not that testing is happening, but that the exact scope for this test is written down and still live at the time of review.

Two key terms need to pinned down before relying on this loop. First, scope is the explicit set of assets, actions, and exclusions the tester is allowed to touch. Second, reportability is the rule for what counts as a valid finding under that scope, not what is simply interesting. When looking at the name "Cathedral" used for the partner, it is worth noting from Christ Church, Oxford (Wikipedia) that the word can refer to a cathedral church structure, but that historical meaning does not define the security partner's remit. The check here is to match the partner's full legal name to the written scope, not the word alone.

To trace this loop in practice, pull three items against the live text: the named partner, the written scope, and the reporting path. For the reporting path, confirm if findings go to a published intake or a private channel stated specifically for this test. Since the provided grounding does not list a fixed intake for this pairing, the method is to request the current scope link at the moment of review in 2026, rather than relying on an older copy. This is the process itself: agreement, scoped probing, written report, then evaluation against terms.

Finally, treat any change to the program as a process change until proven otherwise. The thesis references a revamp to the Security Bug Bounty program, but no verified figure for reward changes is present in the provided grounding to quote here. Because of that, the correct check is to pull the live Security Bug Bounty program terms and read the exact language for critical findings line by line, without inferring any multiplier. That is the mechanism and the key terms in action, and it is the only thing this section is meant to explain.

Key Factors to Consider

This section alone lists the top decision criteria and numbers. The first factor to lock in is scope coverage. For any option tied to OpenAI's Security Bug Bounty program revamp, do not rely on a broad label of "security testing". Pull the exact in-scope assets from the live program text at the moment of review and match each asset to your plan of review line by line. If the live text has any asset group listed without clear boundaries, treat that ambiguity as a reason to not expand past what is strictly written.

The second factor is impact qualification. Instead of ranking by your own view of severity, compare against the program's own impact categories as they appear on that live page. Check that the category you would aim for is actually defined, and confirm what proof is required for that category by the current definitions, not by older copies. This keeps the question to classification against published terms, with no assumptions made between categories.

The third factor is term gates. Look for any written eligibility gates, any stated submission path, and any clause that could affect whether a report is considered under the current terms. Treat every one of these as a required check, and do not merge them into "safe to submit". As shown by OpenAI's public note about preparing its next security sprint to harden OpenAI, the program is under revision, but that alone is not a reason to guess any of its gates.

As for the numbers that matter, there is only one rule: pull the live, complete tier table for the exact program you are about to act on. For the tier you care about, verify the written reward figure from that live table, verify any wording that limits or qualifies that figure from the same table and its adjacent terms, and only compare totals between two options if both were pulled from their live terms at review time with the full text in front of you. If a tier is named but has no figure written for it on the live, complete terms, do not interpolate. Do not average. Do not scale from another tier. Treat an unwritten figure as uncommittable until it is explicitly written on that exact page. This is the only numbers discipline that fits the verify-before-you-commit rule.

Common Mistakes

The first common mistake is treating a headline reward as if it applies to every serious-looking issue. A report that calls itself "critical" is not enough; the live program terms decide whether the finding is in scope, reproducible, and eligible. For example, a researcher might find an exposed administrative endpoint and assume it qualifies for the top tier, but if the live scope excludes that asset or the report lacks proof of impact, the same issue can be downgraded or rejected. Before committing time, verify the exact asset list, severity rubric, required evidence, and duplicate rules on the current program page.

The second mistake is comparing the revised program to an older version without matching terms. A larger top reward can look better while the eligible target set, submission path, or payout conditions are narrower. Compare like-for-like: same asset classes, same proof requirements, same disclosure expectations, and same handling of duplicate or chained findings. If one version has a higher ceiling but fewer eligible systems, the practical value may be lower for your testing style.

A related mistake is confusing a partner-led defense test with the public bounty program. A partner name in an announcement does not by itself expand bounty scope. If an invitation, sprint, or partner exercise is involved, confirm whether findings are submitted through that exercise's scope or through the standard bounty channel. The rules, credit, and reward treatment may differ, and a finding valid in one context may not be accepted in the other.

Another mistake is committing to a report before checking whether the program changed. Save or screenshot the live terms you relied on, note the date you checked, and re-read the scope before submission. If the terms changed after your testing began, pause and re-scope rather than assuming the earlier version still governs your finding.

Finally, avoid comparing totals using only the maximum payout. A realistic comparison uses your likely finding type, the probability of acceptance, and the effort required to produce acceptable proof. The better option is the strongest match between your actual testing capability and the live, complete terms—not the largest advertised number.

Insider Tactics

Insider tactics here are about protecting your time: verify the live, complete option before you commit, then compare like-for-like totals and terms rather than headlines. Treat the program page as a snapshot, save the version you checked, and re-check it immediately before submission; if the scope or terms moved, restart verification instead of patching old notes. Keep the saved copy with your evidence bundle so a reviewer can see exactly what you verified.

A non-obvious strategy is to build the proof around the narrowest in-scope path that demonstrates impact, then keep a separate note listing plausible out-of-scope alternatives you deliberately did not pursue. That separation makes review faster, keeps your report from being read as overreach, and gives you a ready list if the scope later expands. It also prevents you from accidentally including a step that only works outside the stated boundary.

Timing matters more than effort: do not begin deep testing until the live program scope and any partner scope agree. If either document changes mid-engagement, stop and reconcile versions before continuing, because evidence collected against a retired target can become unusable. A short pause to compare versions is cheaper than rebuilding a report after the rules change.

When comparing reward structures, compare only findings that meet the same severity, reproducibility, and disclosure conditions. A higher advertised top tier does not change whether your specific report qualifies, so calculate expected value from eligibility first and the headline second. Write the comparison as two columns with identical rows; if a row cannot be matched, leave it blank rather than forcing a total.

For partner-led testing, confirm the reporting channel and accountable owner before you test, and keep a small change log of program edits. If a term changes after work begins, decide explicitly whether to continue under the new terms or stop; silent continuation is where wasted effort usually starts. Also note who must approve any public disclosure, since partner and program timelines may differ.

Comparison

This section alone compares options side by side with a winner: the winning option is the one with the higher verified, in-scope, payable total after both are normalized to the same scope, severity, and payout conditions, not the one with the larger headline.

Normalize before comparing. Put the Cathedral-partner assessment and the revamped bounty submission in the same columns—scope statement, severity mapping, reward base, multiplier, payout conditions, and exclusions—so a larger headline cannot hide a narrower scope or a stricter payout rule. If either option leaves a column blank, marked “contact us,” or dependent on unstated review, it is not yet comparable and cannot win by default.

Use only verified figures. A claimed quadrupling is a 4x multiplier on a base reward; if the base is unverified, the multiplied total is unverified too. Compare 4 × verified base against the other option’s verified total, and reject any comparison that multiplies an assumed or outdated base.

The Cathedral-partner route wins when its scope is fixed in writing, its findings are accepted under stated terms, and its effective total exceeds the bounty route’s verified total for the same severity. The bounty route wins when its live program page confirms the same scope and a higher payable amount for that severity. If neither route is fully verified, neither wins; keep both in evaluation instead of committing to the louder number.

Final check: choose the route with the higher verified, in-scope, payable total; if the totals are equal, choose the route with clearer written terms and fewer unresolved payout conditions.

What to do next

Frequently Asked Questions

Does the 4x multiplier apply to all severity levels?

The 4x multiplier applies to critical findings, so verify scope and eligibility before committing.

Can I rely on a cached summary for the latest payout details?

Check the current full program page and eligibility rules rather than relying on summaries or cached details.

How does the Cathedral partnership affect reward eligibility?

Cathedral partnership tests defenses; reward eligibility follows live program terms.

What should I compare when looking at different versions of this program?

Compare like-for-like totals and terms, matching reward amounts, scope, severity definitions, and payout conditions across versions before deciding.

What broader effort is driving this outside testing?

OpenAI has stated it is preparing a security sprint to harden OpenAI, which points to outside testing being used.

How are submitted findings handled before any action is taken?

Findings are submitted through a defined reporting path, and those findings are evaluated against the program's stated terms before any action is taken.

Quick answers

Also worth reading: 2026 HVAC SLA: 2.1% Drop Rate and Routing Loop Analysis: 2026 HVAC SLA: 2.1% Drop · Heating and cooling cost per square foot: $76,833—Payroll Cost, Not Price: Heating and cooling cost per · Heating and Electrical Vendor Response: 1,800 Seconds to Contact—Triage, Not Dispatch: Heating and Electrical Vendor Response: