Skip to content
NLEN
Illustration: OpenAI pipeline found 948 Linux root bugs in one run: what autonomous bug hunting means

Codex 0.169: codex exec no longer loads project config in an untrusted folder

By Ivo Donker · — text: Claude (AI)

What changes in Codex rust-v0.169.0? The most important change is a security fix: codex exec keeps a folder's project configuration and MCP servers disabled until you trust that folder, even with the sandbox set to workspace-write. If you do want them, you now say so explicitly with --trust. The terminal also gets multiple side chats in tabs, and you can add your own model providers. The release has been on GitHub since November 13, 2026 and is the current stable version.

We read the release notes for rust-v0.169.0 on GitHub and the list of merged changes since rust-v0.168.0.

What happened

OpenAI published Codex rust-v0.169.0 on November 13, 2026 at 08:37 UTC. It is a stable release; only alpha builds of 0.171 followed. On November 15, 0.169.0 was still the newest stable tag.

The change with the biggest consequences sits under bug fixes. Up to this version, codex exec, the non-interactive mode that often runs in scripts and CI, could load a folder's project configuration and MCP servers even when you had not marked that folder as trusted. A repository you just cloned can carry such a configuration. From 0.169.0 they stay off, also with --sandbox workspace-write. Passing --trust to codex exec or codex review explicitly saves trust for that folder.

New features

Other fixes

The Yes, clear context and implement button now hands the full approved plan to the fresh session and keeps the current one open if the handoff fails. On Windows the sandbox handles quoted cmd.exe paths and PowerShell sessions with an unreadable parent folder. The terminal no longer waits on MCP servers a shared app server did not report, and interrupted turns close cleanly. The bundled SQLite moves to 3.54 for write-ahead log fixes.

What this means

If you run Codex in scripts or pipelines, the --trust change is the reason to update, and also the reason to review those scripts. A job that silently used an MCP server from the project folder will not do so in 0.169.0. That is intended: a configuration file in someone else's repository is also a way to make an agent start a tool. We explained earlier how prompt injection through files and knowledge bases works; an MCP server loaded by accident is a related risk.

The direction matches what other makers of AI agents are doing: trust becomes explicit per folder or project. Claude Code earlier tightened its handling of permissions and secrets in a similar way. More background on the risks is in our overview of AI and cybersecurity.

Side chats and collapsing turns are mostly convenience. Custom providers matter more for teams that want to route Codex through their own gateway or another model; that no longer needs hand-edited configuration files.

What we did not check

We did not install 0.169.0 ourselves or reproduce the --trust change. The release notes name no CVE and no affected versions; we do not know whether the old behaviour was exploited. We did not find an official OpenAI post on X about this release; a changelog account on X summarized it.

Checked on November 15, 2026.

This article was written with AI (Claude) and checked against the sources below.

Sources