September 10 Update — Persistent Suspicious Activity Banner + New GPT-5.6 Sol High/Extra High Long-Task Failures, Connection Interruptions and Message Delivery Timeouts
I am respectfully posting another update because the original suspicious-activity problem has still not been fully resolved, and since September 9 I have also begun experiencing a new cluster of serious reliability problems with GPT-5.6 Sol High and Extra High in regular Chat and Project conversations.
I would greatly appreciate it if an OpenAI staff member who sees this could pass the complete issue cluster to the appropriate Account Security, ChatGPT reliability, model-routing, conversation-runtime, session-management, or engineering teams.
I want to document everything clearly because these problems are now preventing me from completing the long, complex workflows that I pay for ChatGPT Pro to perform.
1. The “Suspicious activity detected” banner is still occurring
The original problem began suddenly on September 7, 2026, at approximately 3:30 PM Eastern Time.
ChatGPT began repeatedly displaying:
“Suspicious activity detected. It looks like someone else may be using your ChatGPT account. Please secure your account to regain access to all features.”
My account is secured.
My account has never been shared with anyone.
I am the only person who uses or has access to my ChatGPT account.
I use ChatGPT only on:
- My personal iPad
- My personal phone
Both devices belong to me and are under my control.
I have already:
- Secured the Google account connected to ChatGPT
- Changed my Google password
- Enabled Google 2-Step Verification
- Confirmed that I control my authentication methods
- Reviewed Active Sessions
- Confirmed that the devices/sessions I recognize are mine
- Tried private/incognito browsing
- Tried multiple browsers
- Tried my other personal device
- Cleared cookies/cache
- Signed out and back in
- Used my normal trusted network
- Confirmed I am not using a VPN
- Confirmed I am not using a proxy
- Confirmed I am not using iCloud Private Relay
- Confirmed I am not using third-party ChatGPT clients
- Confirmed I am not using account-sharing services
- Confirmed I am not using automation software to share or distribute account access
The banner has appeared after completely ordinary questions, simple greetings, new chats, existing chats, file work, and long-form writing tasks.
Sometimes it disappears temporarily. Then it comes back.
That means the alert is intermittent, but it is not reliably resolved.
OpenAI’s own Suspicious Activity Alert documentation says these alerts can be cautionary and do not necessarily indicate wrongdoing. OpenAI also explicitly permits one individual to use their account on multiple devices. (OpenAI Help Center)
2. “Log out of all devices” still cannot be completed
One of my legitimate device entries has previously displayed:
“Session status unavailable.”
That device belongs to me.
OpenAI’s Active Sessions documentation specifically says that when a session shows “Session status unavailable,” the user should use Log out of all sessions.
I have tried exactly that.
I select:
Settings → Security → Active Sessions → Log out of all sessions → Log out of all devices
and ChatGPT returns:
“We were unable to process your request at this time. Please try again later.”
I have screenshots documenting this.
So I cannot even successfully execute the session-reset function that OpenAI’s own security documentation tells me to use. OpenAI says Log out of all sessions should terminate the current session and other active sessions, normally within up to 30 minutes. (OpenAI Help Center)
I have exhausted the reasonable security troubleshooting available to me.
3. A new reliability problem began around midday September 9
For approximately three days before this new problem, ChatGPT had actually been functioning much better for my long complex work, aside from the suspicious-activity banner.
Then, around approximately noon Eastern Time on September 9, my GPT-5.6 Sol High/Extra High long-complex-task workflow began failing again.
The failure pattern is very specific.
I upload the required files.
ChatGPT recognizes the files.
I submit my long, detailed prompt.
GPT-5.6 Sol High or Extra High begins doing the requested task normally.
The activity view shows it:
- inspecting story materials
- reading continuity files
- opening predecessor material
- using file tools
- searching the web
- processing lore
- analyzing images
- reconstructing continuity
- planning or beginning the requested continuation
In other words, the task is actually underway.
Then, after substantial processing—often around the half-hour range in my recent attempts—the run freezes or stops progressing.
I then receive:
“Connection interrupted. Waiting for the complete answer.”
In other attempts, after all of that work has already occurred, the run ultimately fails with:
“Message delivery timed out. Please try again.”
I have screenshots showing both states.
This is not simply a prompt failing immediately.
The model has already spent substantial time reading files, researching, analyzing context, and beginning the task before the connection/execution failure occurs.
That is particularly damaging for long-context workflows because the user can spend roughly half an hour waiting while ChatGPT visibly performs substantial work, only to receive no completed output.
4. Simple questions continue to work, while the complex workflows I pay for are failing
I can still ask ChatGPT a simple question.
I can still send ordinary messages.
The platform is not completely offline.
That is not the issue.
I am a ChatGPT Pro subscriber because I use the advanced capabilities for demanding work: long-context reasoning, large uploaded files, research, continuity analysis, complex worldbuilding, Project conversations, and long-form generation.
I am not paying for Pro merely so that I can ask short questions that almost any basic chat interface can answer.
The advanced workflows are the reason I subscribe.
Those are precisely the workflows that are now unreliable again.
My ongoing long, complex worldbuilding stories require ChatGPT to read extensive continuity material, maintain state across large files, research current lore when required, reason through complicated instructions, and then actually finish a substantial response.
Right now, those tasks are being interrupted before completion.
That has forced me to halt my story work and other complex tasks.
5. Regular Chat/Project conversations and Work do not appear to be behaving the same way
Another detail that I think engineering should investigate is that, in my testing, Work has remained substantially more reliable for complex execution while ordinary Chat and Project conversations are experiencing these failures.
I am not saying that proves what the internal architecture is doing.
I do not have access to OpenAI’s backend.
But when the same account, network, and user can perform complex work more successfully through one ChatGPT surface while another surface repeatedly freezes, interrupts, or times out, that seems like useful diagnostic evidence.
There are also recent community reports describing similar differences between regular Chat and Work/Codex, including long-running regular Chat execution ending around the 23–26 minute range. I am mentioning those only as corroborating community reports, not as proof of a specific root cause. (OpenAI Developer Community)
6. OpenAI also acknowledged increased Pro/Plus conversation errors on September 9
OpenAI’s public Status page reported an “Increased Error Rate For Pro and Plus Plan Conversations” incident on September 9.
The incident was investigated, mitigated, and later marked resolved. (OpenAI Status)
I appreciate that acknowledgment because it confirms that there really were conversation-level problems affecting paid customers during the same day that my new failures began.
However, I am still documenting my own failures because my symptoms involve long-running GPT-5.6 Sol High/Extra High execution, file/tool processing, connection interruption, and message-delivery failure, and I do not know whether the resolved public incident fully covers those symptoms.
If it does, I would appreciate confirmation.
If it does not, I would appreciate engineering treating these reports separately rather than assuming that a green/resolved status automatically means all affected workflows are now healthy.
7. I have noticed a recurring pattern around major model transitions, but I am not claiming causation
I also want to raise one observation carefully.
GPT-6 Astra began rolling out on September 3. OpenAI described it as a phased rollout over the following days. (OpenAI)
In my personal experience using ChatGPT over time, periods surrounding major model launches or model retirements have sometimes coincided with instability in older/current models: reasoning freezes, abnormal execution, connection interruptions, UI problems, file reliability issues, or changes in how long-running tasks behave.
I remember experiencing similar instability around the GPT-5.6 launch period, when GPT-5.5-related workflows and ChatGPT reliability were also behaving inconsistently for me.
I want to be very explicit:
I am not claiming that GPT-6 Astra caused the current GPT-5.6 problems.
I cannot see OpenAI’s infrastructure and I do not know the root cause.
This is only a hypothesis based on timing and previous patterns I personally experienced.
What I am respectfully asking is whether engineering can investigate whether any rollout-related changes involving routing, worker allocation, capacity, queues, session execution, model serving, long-running-task handoff, or shared infrastructure could be contributing to the differences users are seeing between regular Chat and Work.
If Astra is completely unrelated, that clarification would also be useful.
8. These issues should not be collapsed into one unsupported explanation
There may actually be multiple problems here:
- the persistent/intermittent suspicious-activity security banner
- the “Session status unavailable” state
- the broken Log out of all devices action
- GPT-5.6 Sol High/Extra High long-task execution failures
- “Connection interrupted. Waiting for the complete answer”
- “Message delivery timed out. Please try again”
- potential differences between regular Chat/Project conversations and Work
- broader Pro/Plus conversation reliability problems
I am not claiming they all have the same root cause.
I am asking OpenAI to investigate them rather than requiring customers to guess.
9. Why this matters to me as a paying Pro user
I understand that operating a service at ChatGPT’s scale is complicated.
I understand that new models and infrastructure changes require enormous work behind the scenes.
I do not expect technology to be perfect.
What is difficult as a paying customer is the uncertainty and silence when a workflow that was functioning properly suddenly becomes unreliable again.
My long-form story and worldbuilding work is not a trivial one-message task.
It depends on stable long-context processing, file access, reasoning, research, and response delivery.
When ChatGPT works for 20–30 minutes, successfully reads everything, performs the research, starts constructing the answer, and then fails with a connection interruption or delivery timeout, I lose the entire productive run.
That makes it impossible for me to trust the platform with the complex tasks for which I am paying.
10. What I am respectfully asking OpenAI staff to do
If @OpenAI_Support or another OpenAI staff member sees this update, I would sincerely appreciate the following:
Please escalate the evidence to the teams responsible for regular Chat/Project execution, GPT-5.6 Sol High/Extra High reasoning, long-running task execution, file/tool state propagation, response streaming/message delivery, session management, account security, and model routing/capacity.
Please investigate why long tasks can successfully ingest files and perform substantial processing before freezing or losing response delivery.
Please investigate the recurring approximate long-run failure point.
Please investigate why Work can remain substantially more reliable while regular Chat/Project workflows fail.
Please investigate whether the September 9 Pro/Plus incident fully covers these failures or whether a separate issue remains.
Please investigate whether any recent model rollout or infrastructure change is relevant—but please understand that I am asking this as a question, not claiming Astra is the cause.
Please continue investigating the persistent suspicious-activity/session problem separately as well.
Most of all, I am respectfully asking for acknowledgment if OpenAI is aware of these problems.
Paying users should not have to repeatedly reproduce serious failures and wonder whether engineering even knows they are happening.
I am happy to provide screenshots, timestamps, shared-chat examples, and other diagnostic evidence.
If other users are experiencing the same combination—especially GPT-5.6 Sol High/Extra High successfully starting complex file-based tasks and then ending with connection interruption or message-delivery timeout—please feel free to add your own evidence below.
I am not trying to attack OpenAI or its staff.
I want the platform to work reliably again so I can continue the long, complex work I use ChatGPT Pro for.
Thank you for reading this and for passing it to the appropriate team if possible.
