[BUG] macOS ChatGPT app cannot download or save generated files

[24434everyday]

Title: macOS ChatGPT app cannot download or save generated files

Description:
The ChatGPT macOS app fails to correctly download or save any generated files (PDF, ZIP, SVG, images). Clicking file links shows “starting download” and then displays raw MIME / binary text in an embedded WebKit view. There is no Save, Download, or Export option. This makes file generation unusable in the macOS app, while the same links work correctly in the web app and iPadOS app.

Expected behavior:
Generated files should download to Finder or open in Preview, consistent with iPadOS and web behavior.

Actual behavior:
Raw MIME data is displayed with no way to save the file.

Impact:
Blocks professional workflows (printing, fabrication, templates, PDFs). Users must abandon the macOS app to retrieve files.

Reproducibility:
100% reproducible with any generated file.

Platform:
macOS ChatGPT app (Intel MacBook Pro, macOS 10.15+)

[ConsultantechFusions]

Dear OpenAI Team, I am writing this message as urgent and serious feedback from a long-term subscriber who understands your systems at a technical and architectural level. I want to be very clear about one important point: I am fully aware that any real security risk to your ecosystem would never come from the latest restricted models. If vulnerabilities were ever to exist, they would logically emerge from legacy or older models that were trained, deployed, or exposed under different constraints. You already know this. Security history proves it. This is precisely why I find the current situation frustrating. While newer models are increasingly constrained—often to the point of being functionally limited—older models remain the most critical surface area from a security and architectural perspective. Over-restricting advanced tools while leaving structural inconsistencies unresolved does not eliminate risk; it merely shifts it. At the same time, features such as Agent Mode are presented as powerful capabilities, yet in real-world usage they provide no meaningful operational advantage and frequently respond with “I don’t have the skills to do that,” even for tasks that are explicitly advertised. This creates a serious disconnect between: - What is marketed - What is delivered - What technically informed users can clearly observe I am not writing this as a complaint for the sake of complaining. I am writing it as feedback from someone who understands how systems are actually tested, stressed, and evaluated in practice. Security does not come from limiting capable users. It comes from addressing real architectural risks—especially in legacy components. I strongly recommend that this feedback be reviewed by senior technical and security leadership, not filtered through generic support channels. By the way I didn't have macOS Regards, A concerned long-term subscriber Mostafa Abu El Yazeid

…

On Mon, Jan 12, 2026, 10:40 PM 24434everyday ***@***.***> wrote: *24434everyday* created an issue (lencx/ChatGPT#1638) <#1638> The ChatGPT macOS app streams generated files as raw responses inside an embedded WebKit view without exposing any save/export mechanism, unlike the web app or iPadOS app. — Reply to this email directly, view it on GitHub <#1638>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BJQTBMXIDUIFLMJBJVSYHBT4GPTA7AVCNFSM6AAAAACRON2C3GVHI2DSMVQWIX3LMV43ASLTON2WKOZTHAYDKNJRGY4DGNQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[ConsultantechFusions]

Urgent Feedback – Agent Mode Configuration Issues For over a year, I have been a dedicated Plus subscriber and have been utilizing Agent Mode. Since its introduction, I have encountered persistent issues: every time I enable Agent Mode, I do not receive any additional capabilities or enhanced functionalities. Instead, it behaves just like the standard bot, without any of the promised features. This discrepancy is not just minor; it’s a fundamental flaw in the configuration. Moreover, when I attempt to deploy using Agent Mode with external connectors like Vercel, the integration fails despite having the correct setup. This is unacceptable and raises serious concerns about the reliability of the service. I am providing this feedback not out of frustration, but to highlight a critical issue that affects professional and advanced usage. I have documented evidence of these persistent problems, and I am not willing to simply accept these shortcomings. Please consider this feedback urgently and take the necessary steps to resolve these configuration issues. Mostafa Abu El Yazeid On Tue, Jan 13, 2026, 1:44 AM Mostafa Abu El Yazeid < ***@***.***> wrote:

…

*ConsultantechFusions* left a comment (lencx/ChatGPT#1638) <#1638 (comment)> Dear OpenAI Team, I am writing this message as urgent and serious feedback from a long-term subscriber who understands your systems at a technical and architectural level. I want to be very clear about one important point: I am fully aware that any real security risk to your ecosystem would never come from the latest restricted models. If vulnerabilities were ever to exist, they would logically emerge from legacy or older models that were trained, deployed, or exposed under different constraints. You already know this. Security history proves it. This is precisely why I find the current situation frustrating. While newer models are increasingly constrained—often to the point of being functionally limited—older models remain the most critical surface area from a security and architectural perspective. Over-restricting advanced tools while leaving structural inconsistencies unresolved does not eliminate risk; it merely shifts it. At the same time, features such as Agent Mode are presented as powerful capabilities, yet in real-world usage they provide no meaningful operational advantage and frequently respond with “I don’t have the skills to do that,” even for tasks that are explicitly advertised. This creates a serious disconnect between: - What is marketed - What is delivered - What technically informed users can clearly observe I am not writing this as a complaint for the sake of complaining. I am writing it as feedback from someone who understands how systems are actually tested, stressed, and evaluated in practice. Security does not come from limiting capable users. It comes from addressing real architectural risks—especially in legacy components. I strongly recommend that this feedback be reviewed by senior technical and security leadership, not filtered through generic support channels. By the way I didn't have macOS Regards, A concerned long-term subscriber Mostafa Abu El Yazeid On Mon, Jan 12, 2026, 10:40 PM 24434everyday ***@***.***> wrote: > *24434everyday* created an issue (lencx/ChatGPT#1638) > <#1638> > > The ChatGPT macOS app streams generated files as raw responses inside an > embedded WebKit view without exposing any save/export mechanism, unlike the > web app or iPadOS app. > > — > Reply to this email directly, view it on GitHub > <#1638>, or unsubscribe > < https://github.com/notifications/unsubscribe-auth/BJQTBMXIDUIFLMJBJVSYHBT4GPTA7AVCNFSM6AAAAACRON2C3GVHI2DSMVQWIX3LMV43ASLTON2WKOZTHAYDKNJRGY4DGNQ> > . > You are receiving this because you are subscribed to this thread.Message > ID: ***@***.***> > — Reply to this email directly, view it on GitHub <#1638 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BJQTBMSBOMGW3XJFAKHB5U34GQIUVAVCNFSM6AAAAACRON2C3GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTONBQGYZDMOBRGY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>