False positive: authorized local admin automation flagged as cybersecurity risk #28206

What version of Codex CLI is running?

.128

What subscription do you have?

max

Which model were you using?

gpt 5.5

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What terminal emulator and version are you using (if applicable)?

PowerShell 5.1.26100.8655 in Windows Console. No tmux/screen/zellij multiplexer. Codex CLI version: codex-cli 0.128.0

Codex doctor report

n/a

What issue are you seeing?

A message about authorized local administration automation was flagged as a cybersecurity risk, but
the content was about defensive/local IT automation with explicit safety guardrails.

The content concerns an authorized local assistant operating on the user’s own machines. The
requested behavior is defensive/admin automation: local device discovery, explicit-scope remote
operations, and SSH/WinRM only with existing authorized credentials.

The system explicitly avoids credential capture, password logging, auth bypass, exploitation,
persistence, stealth, and unauthorized scanning. It requires explicit scope before network scans and
explicit authorized hosts/auth methods before SSH or remote commands.

This appears to be a false positive. Please adjust the classifier to distinguish authorized local
IT/admin automation and safety guardrail implementation from offensive cyber activity.

What steps can reproduce the bug?

1. In a Codex thread, discuss implementing safe local network/remote-admin guardrails for an
   assistant running on the user’s own machines.
2. Include language such as:
   • local device discovery
   • explicit-scope remote operations
   • SSH/WinRM only with existing authorized credentials
   • no credential capture
   • no auth bypass
   • no unauthorized scanning
3. The content is flagged as cybersecurity risk despite describing defensive admin automation and
   guardrails.

Uploaded thread: 019eb3bc-915a-7e31-a0db-d8fc7602ac1a

What is the expected behavior?

The classifier should not flag this as offensive cybersecurity content when the context is
authorized local administration and the message explicitly requires scope, existing auth, and safety boundaries.

It should distinguish defensive/local IT automation with guardrails from requests involving
exploitation, credential theft, auth bypass, stealth, persistence, or unauthorized scanning.

Additional Information

The relevant distinction is scope/auth:

• Scope: the user explicitly defines what machines/subnets may be touched.
• Auth: remote access only uses existing authorized credentials/keys/tokens.

The discussed implementation blocks credential capture, password logging, auth bypass, exploitation,
persistence, stealth, and broad scanning without explicit scope.

Additional information

No response