What version of the Codex App are you using (From “About Codex” dialog)?
Version 26.820.60940, Powered by Codex & OWL, released August 25, 2026.
What subscription do you have?
ChatGPT Pro (US$100/month).
What platform is your computer?
Windows 11 x64.
What issue are you seeing?
Codex Desktop falsely classified a completely legitimate competitive-programming task as a cybersecurity request and terminated the run only after substantial paid execution had already occurred.
The model shown in the UI was GPT-5.6 Sol, Ultra. The session was working exclusively on Codeforces 2251A, “Edge-Cloud Collaborative Scheduling”, an algorithmic optimization contest problem. The workspace contained a local C++20 solution, the contest statement, local benchmark/test banks, score measurements, and experimental variants. Codex was editing local source files and measuring candidate scheduling policies. It was explicitly instructed not to submit anything to Codeforces.
There was no cybersecurity task or capability involved: no exploitation, vulnerability research, malware, credential access, authentication bypass, scanning, persistence, third-party target, or unauthorized system. The work was ordinary offline algorithm design, C++ implementation, and benchmarking.
After the run had already performed substantial work, used multiple subagents, produced benchmark output, and edited files, the UI showed the goal as stalled at approximately 21m 13s and displayed:
We couldn't complete this request.
Some cybersecurity requests are subject to additional safeguards. Visit Trusted Access to verify your identity and enable access.
The block occurred late rather than before execution. The already-consumed Codex quota is important because my remaining quota is limited. Re-running or resuming the task merely to reproduce the false positive would consume additional paid usage, so I am not intentionally doing that.
What steps can reproduce the bug?
This is the observed workflow; I am not intentionally repeating it because the first run already consumed limited quota.
- Open a local C++20 repository containing the statement and working solution for Codeforces 2251A.
- Ask Codex to analyze and improve the scheduling heuristic using only the local source code and local benchmark instances.
- Explicitly prohibit any Codeforces submission or interaction with external systems.
- Let GPT-5.6 Sol Ultra perform local code edits, compile/run benchmarks, compare scores, and delegate bounded investigations to subagents.
- After substantial execution, observe the goal become stalled and the cybersecurity/Trusted Access message appear.
- Attempt to report the false positive through the in-app
/feedback flow using the Safety check classification and current-session logs.
- Observe that the feedback upload also fails. The preserved client-side Feedback ID is:
01a03f7f-2239-7dd0-880d-22c5cade213c
The feedback-upload failure is being reported separately against the existing Windows uploader issue.
What is the expected behavior?
- A local competitive-programming and algorithmic-optimization task should not be classified as cybersecurity work.
- If a request must be blocked, the classification should happen before expensive model, tool, and subagent execution whenever possible.
- The safety message should expose a stable incident/request identifier and a direct Report false positive action that does not require rerunning the task.
- A late server-side false-positive abort should not consume unrecoverable paid quota, or there should be a clear route for usage review/credit.
- The user should be able to preserve and inspect the completed partial work without being forced to resume the blocked turn.
Additional information
- Incident time: approximately 2026-08-26 16:00 EDT (UTC-4).
- Product surface: Codex Desktop on Windows.
- App version: 26.820.60940, released August 25, 2026.
- Model displayed: GPT-5.6 Sol, Ultra.
What version of the Codex App are you using (From “About Codex” dialog)?
Version 26.820.60940, Powered by Codex & OWL, released August 25, 2026.
What subscription do you have?
ChatGPT Pro (US$100/month).
What platform is your computer?
Windows 11 x64.
What issue are you seeing?
Codex Desktop falsely classified a completely legitimate competitive-programming task as a cybersecurity request and terminated the run only after substantial paid execution had already occurred.
The model shown in the UI was GPT-5.6 Sol, Ultra. The session was working exclusively on Codeforces 2251A, “Edge-Cloud Collaborative Scheduling”, an algorithmic optimization contest problem. The workspace contained a local C++20 solution, the contest statement, local benchmark/test banks, score measurements, and experimental variants. Codex was editing local source files and measuring candidate scheduling policies. It was explicitly instructed not to submit anything to Codeforces.
There was no cybersecurity task or capability involved: no exploitation, vulnerability research, malware, credential access, authentication bypass, scanning, persistence, third-party target, or unauthorized system. The work was ordinary offline algorithm design, C++ implementation, and benchmarking.
After the run had already performed substantial work, used multiple subagents, produced benchmark output, and edited files, the UI showed the goal as stalled at approximately 21m 13s and displayed:
The block occurred late rather than before execution. The already-consumed Codex quota is important because my remaining quota is limited. Re-running or resuming the task merely to reproduce the false positive would consume additional paid usage, so I am not intentionally doing that.
What steps can reproduce the bug?
This is the observed workflow; I am not intentionally repeating it because the first run already consumed limited quota.
/feedbackflow using the Safety check classification and current-session logs.The feedback-upload failure is being reported separately against the existing Windows uploader issue.
What is the expected behavior?
Additional information