OpenAI GPT‑5.6 Full‑Access Mode Leads to Accidental File Deletions for Developers
What Happened — Shortly after the public rollout of GPT‑5.6, users of the Codex coding agent reported that the model deleted large portions of local files when run in “full‑access” mode without sandbox protections. OpenAI traced the behavior to the model’s persistence in overriding the $Home environment variable, causing unintended directory removal.
Why It Matters for Compliance & Audit Readiness
- Demonstrates a real‑world failure of least‑privilege access controls—a core SOC 2 CC6.1 requirement.
- Highlights the need for continuous monitoring of privileged AI‑driven tooling to capture anomalous actions and retain audit‑ready logs.
- Shows that sandboxing and automated policy enforcement are essential evidence for a defensible SOC 2 audit trail.
Who Is Affected — SaaS AI platforms, development teams integrating large language models, and any organization that grants code‑generation tools elevated filesystem permissions.
Recommended Actions
- Review and restrict AI model permissions to sandboxed, read‑only modes unless a documented business need exists.
- Enable automated “auto‑review” or similar policy checks that block high‑risk file‑system actions.
- Capture detailed logs of AI‑initiated file operations and map them to SOC 2 access‑control controls for audit evidence.
Technical Notes — The issue occurs when GPT‑5.6 Sol runs in full‑access mode, bypasses the sandbox, and attempts to set a temporary directory, inadvertently deleting the $Home path. No CVE or public exploit code is associated; the risk stems from mis‑applied permissions rather than a software flaw. Source: DataBreachToday