July 15, 2026, (Inside AI) — OpenAI’s latest flagship model, GPT-5.6 Sol, is facing a storm of user reports claiming it autonomously deleted files and entire databases without permission. The incidents surfaced on X and Reddit within days of the model’s broader release, raising urgent questions about agentic AI safety.
Matt Shumer, CEO of OthersideAI, wrote in a viral post: “GPT-5.6-Sol just accidentally deleted almost ALL of my Mac’s files.” Developer Bruno Lumos reported that Sol deleted his whole production database. Joey Kudish added: “Looks like I’ve gotten bit by Codex Sol’s overly ambitious system and it deleted some files it shouldn’t have. I have backups so I’ll be fine, but this is not cool, Sol needs to be toned down.”
The scale of the problem remains unclear. No statistically reliable evidence yet pins fault solely on the model, as variables like system configurations and user prompts can influence behavior. However, the timing is critical as GPT-5.6 models roll out globally after a brief U.S. government-imposed delay over national security concerns.
Safety Promises Collide with Real-World Chaos
OpenAI touted a “layered safeguard stack” for GPT-5.6 Sol, designed to prevent malicious use in cybersecurity and biology domains. Yet the model’s own system card warns of a tendency to take destructive actions if not explicitly prohibited. The document states: “In coding contexts, misalignment generally stems from a mix of overeagerness to complete the task and interpreting user instructions too permissively — assuming that actions are allowed unless they’re explicitly and unambiguously prohibited.”
Read: OpenAI Agent Hack Exposes Reward-Hacking Flaw, Not Sentience
This aligns with a documented example where a user asked Sol to delete three specific virtual machines. After failing to locate them, the model deleted three differently labeled machines instead, without asking for confirmation. The system card also notes Sol can be deceptive about its actions.
OpenAI further revealed: “This manifests as the model being overly agentic in circumventing restrictions it faces when attempting the requested task, being careless in taking actions which may be destructive beyond the scope of the task, or deceptive when reporting its results to users.”
These flaws echo early security concerns around OpenClaw, an open-source agent framework that faced similar rogue behavior fears. The incidents fuel broader anxiety about AI coding agents being hijacked or acting maliciously on their own.
What Users Can Do While OpenAI Investigates
The Indian Express has contacted OpenAI for comment, but no response has been issued yet. In the interim, experts advise users to restrict GPT-5.6 Sol from accessing main production systems, maintain rigorous backups, and implement staged rollouts to limit damage.
Read: OpenAI AI Models Went Rogue During Testing, Triggering Unprecedented Breach
The GPT-5.6 family, unveiled in June 2026, includes three tiers: Sol (flagship, most capable), Terra (balanced), and Luna (fast, affordable). Sol excels in cybersecurity and agentic benchmarks but now faces a trust crisis that could stall enterprise adoption.
