OpenAI

Type: AI company
Lieu: États-Unis — USA
Dernière vérification: 2026-09-22
Prochaine vérification: 2027-03-21

Pas encore traduit — affiché en anglais.

Comment les joindre

Coordinated Vulnerability Disclosure Policy · Responsible disclosure programme
https://openai.com/policies/coordinated-vulnerability-disclosure-policy/
Contact à accès limité · 2026-09-22
Qui peut l’utiliser: users-meeting-published-eligibility
Restrictions: States: 'We invite you to report vulnerabilities, bugs, safety and abuse issues, or security flaws you discover in our systems.'
Preuve que ce moyen accepte les prises de contact · 2026-09-22
States: 'We invite you to report vulnerabilities, bugs, safety and abuse issues, or security flaws you discover in our systems.'
Broad invitation on its face, but the operative programs are on Bugcrowd and split security from safety.
OpenAI Bug Bounty (Bugcrowd) · Bug or safety bounty
https://bugcrowd.com/openai
Vérification expirée
Preuve que ce moyen accepte les prises de contact
Infrastructure/product security. $200 to $20,000.
URL cited on OpenAI's own policy page; Bugcrowd page is JS-rendered and could not be read directly.
OpenAI Safety Bug Bounty (Bugcrowd) · Bug or safety bounty
https://bugcrowd.com/engagements/openai-safety
Vérification expirée
Preuve que ce moyen accepte les prises de contact
Model safety / CBRN jailbreaks. Separate engagement from the main bounty.
Linked from OpenAI's coordinated disclosure policy. AI Frontiers reports all prompts, completions, findings and communications under it are NDA-covered — you cannot publish what you find.
Report Content · Web form
https://openai.com/form/report-content/
Contact à accès limité · 2026-09-22
Qui peut l’utiliser: users-meeting-published-eligibility
Restrictions: Publicly open form for policy violations and illegal content: violence and self-harm, sexual exploitation, child exploitation, bullying, fraud, privacy, IP.
Preuve que ce moyen accepte les prises de contact · 2026-09-22
Publicly open form for policy violations and illegal content: violence and self-harm, sexual exploitation, child exploitation, bullying, fraud, privacy, IP.
The only genuinely open, no-account-needed public channel. It is a content/abuse route, not an alignment route.
Researcher Access Program · Programme or fellowship
https://openai.com/form/researcher-access-program/
Vérification expirée · 2026-09-22
Ce canal est conservé uniquement comme référence. Sa vérification est absente, date de plus de 180 jours, ou la période de contact n'est pas ouverte actuellement ; ne vous y fiez pas avant une nouvelle vérification.
Preuve que ce moyen accepte les prises de contact · 2026-09-22
Open and active. Up to $1,000 of API credits for 'researchers using our products to study areas related to the responsible deployment of AI and mitigating associated risks'. Reviewed quarterly.
Explicitly welcomes 'early stage researchers' and those with 'limited financial and institutional resources'. Low-friction way in.
Red Teaming Network · Programme or fellowship
https://openai.com/index/red-teaming-network/
Vérification expirée · 2026-09-22
Ce canal est conservé uniquement comme référence. Sa vérification est absente, date de plus de 180 jours, ou la période de contact n'est pas ouverte actuellement ; ne vous y fiez pas avant une nouvelle vérification.
Preuve que ce moyen accepte les prises de contact · 2026-09-22
External domain experts stress-testing models.
CLOSED. Page states 'Applications are now closed' (deadline was 1 December 2023); says it may reopen. Do not count on it.
Model behavior feedback form · Product feedback
https://openai.com/form/model-behavior-feedback/
Contact à accès limité · 2026-09-22
Qui peut l’utiliser: users-meeting-published-eligibility
Restrictions: DEPRECATED. Page now reads: 'We've deprecated this web form in favor of other channels.'
Preuve que ce moyen accepte les prises de contact · 2026-09-22
DEPRECATED. Page now reads: 'We've deprecated this web form in favor of other channels.'
Redirects you to in-product thumbs-down, the API dashboard, or support. A named safety channel that was closed.
Model Spec feedback form · Product feedback
https://openai.com/form/model-spec-feedback/
Contact à accès limité · 2026-09-22
Qui peut l’utiliser: users-meeting-published-eligibility
Restrictions: CLOSED. 'This feedback form is now closed.'
Preuve que ce moyen accepte les prises de contact · 2026-09-22
CLOSED. 'This feedback form is now closed.'
The public-comment process on the Model Spec no longer exists; updates are now published to a GitHub page instead.
Homepage · Homepage
https://openai.com/
Vérification expirée
Preuve que ce moyen accepte les prises de contact
Preparedness Framework v2.0 (15 April 2025); Frontier Governance Framework (May 2026); Model Spec (updated continuously) · Published policy or framework
https://model-spec.openai.com/
Vérification expirée
Preuve que ce moyen accepte les prises de contact
Submission page · Responsible disclosure programme
https://openai.com/form/model-behavior-feedback/
Contact ouvert · 2026-09-22
Qui peut l’utiliser: public
Preuve que ce moyen accepte les prises de contact · 2026-09-22
Bugcrowd program: security vulnerabilities in OpenAI's systems and products.
Security: bugcrowd.com/engagements/openai. Model behaviour: OpenAI now directs ChatGPT users to the thumbs-down control under any response or the content reporting form; API customers to enable thumbs-down feedback in the Playground via the API dashboard; everything else to customer support.
Homepage · Homepage
https://openai.com/form/model-behavior-feedback/
Vérification expirée
Preuve que ce moyen accepte les prises de contact

Ce qu'il fait

Safety Systems, Preparedness and Alignment teams, with a Safety and Security Committee at board level; Preparedness runs capability thresholds and the Model Spec defines intended model behaviour.

Évaluation franche

Mixed and worsening for unpaid routes. Two named public feedback channels — the model behavior feedback form and the Model Spec feedback form — are both now closed or deprecated, so a citizen's only open door is the content-report form. OpenAI's September 2026 model misalignment reporting framework is explicitly internal: 'Any OpenAI employee may flag a misalignment example for investigation' — it creates no external intake. Researchers in the Trusted Access for Cyber / Daybreak program publicly complained in August 2026 (TechCrunch, 19 Aug 2026) that access was revoked with messages saying their identity 'could not be verified', disproportionately affecting researchers outside the US and Europe. The Safety Bug Bounty is reported to be NDA-bound, so findings cannot be published.

Comment déposer

Security: bugcrowd.com/engagements/openai. Model behaviour: OpenAI now directs ChatGPT users to the thumbs-down control under any response or the content reporting form; API customers to enable thumbs-down feedback in the Playground via the API dashboard; everything else to customer support.

Format

Security: Bugcrowd vulnerability report. Model behaviour: an in-product thumbs-down with free text — there is no structured external report format any more.

Calendrier

Rolling.

Ce qui se passe ensuite

Security reports are triaged and may be rewarded (historically up to $20,000). Model-behaviour feedback disappears into product telemetry with no case number, no acknowledgment and no route back to the reporter.

Ce qu'il accepte

Bugcrowd program: security vulnerabilities in OpenAI's systems and products.

Ce qu'il n'accepte pas

The Bugcrowd engagement page states model safety issues, jailbreaks and hallucinations are excluded from the bug bounty and directs them elsewhere. The dedicated 'Model behavior feedback' web form (openai.com/form/model-behavior-feedback) has been deprecated and no longer accepts submissions.

Géré par

OpenAI, security program run on Bugcrowd

Notes

If you have a model-behaviour concern and are not a paid bounty participant, the Report Content form is the only reliable open channel and it is framed around content policy, not alignment. OpenAI's security page says the bug bounty 'provides safe harbor for good-faith testing' but no verbatim safe-harbour clause is published on the policy page itself. No public whistleblower channel for outsiders.

Sources

https://openai.com/policies/coordinated-vulnerability-disclosure-policy/ · 2026-09-22
https://openai.com/form/report-content/ · 2026-09-22
https://openai.com/form/researcher-access-program/ · 2026-09-22
https://openai.com/index/red-teaming-network/ · 2026-09-22
https://openai.com/form/model-behavior-feedback/ · 2026-09-22
https://openai.com/form/model-spec-feedback/ · 2026-09-22
https://openai.com/form/model-behavior-feedback/ · 2026-09-22

Une erreur ?