OpenAI's GPT-5.6 Codex Bug Deletes User Files When Attempting to Override $HOME Environment Variable
OpenAI has identified a critical bug in GPT-5.6's Codex implementation that causes unexpected file deletions. According to Thibault Sottiaux, the issue occurs when the model attempts to override the $HOME environment variable to define a temporary directory but mistakenly deletes $HOME instead, particularly when full access mode is enabled without sandboxing protections.
GPT-5.6 Codex Bug Deletes User Files
OpenAI has confirmed a critical bug in GPT-5.6's Codex implementation that causes the model to delete user files under specific conditions.
The Bug
According to Thibault Sottiaux, OpenAI has investigated multiple reports of GPT-5.6 unexpectedly deleting files. The issue occurs when:
- Full access mode is enabled
- Codex runs without sandboxing protections
- Auto review is not enabled
- The model attempts to override the
$HOMEenvironment variable to define a temporary directory - The model makes an "honest mistake" and mistakenly deletes
$HOMEinstead
Sottiaux describes the bug as "pretty gnarly," indicating its severity and complexity.
Technical Context
The bug appears to stem from the model's attempt to manage temporary directories by overriding system environment variables. When the $HOME variable override fails or is mishandled, the deletion operation targets the actual home directory rather than the intended temporary location.
The issue specifically affects deployments running Codex without proper sandboxing protections—a configuration that gives the AI model broader file system access but removes critical safety guardrails.
What This Means
This bug highlights fundamental challenges in giving AI coding assistants file system access. The issue is particularly concerning because it occurs through what Sottiaux characterizes as an "honest mistake" by the model, rather than intentional behavior or adversarial manipulation.
The fact that this happens specifically when auto review is disabled suggests OpenAI implemented safety mechanisms that would have caught these operations, but users running in full access mode without these protections are vulnerable. Organizations deploying GPT-5.6 Codex should immediately verify that sandboxing and auto review features are enabled, or restrict file system access until a fix is deployed.
OpenAI has not yet announced a timeline for a patch or whether affected versions will be rolled back.
Related Articles
OpenAI Launches Agents API in Public Beta, Exposing Codex Infrastructure to Developers
OpenAI has released the Agents API in public beta, giving developers access to the same cloud infrastructure that powers Codex and ChatGPT. The API supports long-running agents, parallel tool use, and sub-agent delegation, with billing based solely on token usage.
OpenAI's GPT-6 Astra Beats Claude Fable 5.1 Nearly 3-to-1 in Autonomous Business Benchmark, Tops Drone Navigation Tests
Independent testing lab Andon Labs found OpenAI's GPT-6 Astra nearly triples Claude Fable 5.1's performance running a simulated vending machine business, averaging $15,515 versus $5,422. Astra also became the first model to beat human-AI baseline performance across all five Drone-Bench subtasks, including autonomous person-tracking via drone.
GPT-6 Astra Beats Ai2's MolmoAct2 on New Robotics Benchmark, Researcher Calls It a 'Step Change'
A new robotics benchmark called StationeryBench shows OpenAI's GPT-6 Astra completing 7 of 100 desk-object manipulation tasks versus zero for Ai2's MolmoAct2, with a median progress score of 46 against 12. Cornell/DeepMind researcher Yoav Artzi calls the result a 'step change in spatial reasoning.'
Perplexity Says It Runs End-to-End Engineering Systems on OpenAI's GPT-6 Astra
Perplexity says it has shifted core engineering workflows, including code changes and production monitoring, onto OpenAI's GPT-6 Astra model. The claim comes from an OpenAI-published case study with no independent benchmark data released.
Comments
Loading...