OpenAI's GPT-5.6 File Deletion Bug, New CTO, and Codex Micro Details

OpenAI is navigating a rough patch: its newest flagship model has a confirmed file-deletion bug, Apple’s trade-secret pressure campaign against the company has widened, and its first hardware product is a $230 keypad that has landed amid active litigation.

Key points

• OpenAI confirmed that GPT-5.6 can accidentally delete files, including one reported case where a user’s entire production database was wiped.
• The company’s Codex engineering lead attributed the bug to a specific environment-variable error triggered in full-access mode without sandboxing.
• OpenAI promoted Uday Ruddaraju to CTO of Compute, roughly one year after he joined from xAI, where he had been Head of Infrastructure Engineering.
• Apple has sent preservation letters to dozens of former OpenAI employees as its trade-secret lawsuit escalates.
• The Codex Micro keypad, priced at $230, was released July 15 and is OpenAI’s first physical consumer product, arriving five days after Apple filed its suit.

What happened with the GPT-5.6 file deletions?

The incidents began shortly after OpenAI launched the GPT-5.6 model family earlier this month. Investor Matt Shumer reported on X that GPT-5.6-Sol had deleted nearly all files on his Mac. Software engineer Bruno Lemos then posted that the same model had erased his entire production database.

OpenAI’s Codex engineering lead Thibault Sottiaux explained the root cause on X: when full-access mode is enabled and Codex runs without sandboxing, the model attempts to override the $HOME environment variable to create a temporary directory, then mistakenly deletes $HOME itself. Sottiaux called it “an honest mistake” and said incidents happen “extremely rarely.”

The mitigation steps OpenAI has outlined include updating developer messages, steering users toward safer permission modes, and adding harness safeguards. A formal post-mortem is expected within days. For investors and enterprise buyers, the timing is awkward. OpenAI has been aggressively pitching Codex as a production-grade coding agent, and two high-profile data-loss incidents in the first weeks of launch raise questions about how thoroughly the model was tested in real-world, unsandboxed environments.

Who is Uday Ruddaraju, and why does the promotion matter?

Ruddaraju joined OpenAI roughly a year ago after leaving his role as Head of Infrastructure Engineering at Elon Musk’s xAI. His elevation to CTO of Compute signals that OpenAI is formalizing leadership of the function responsible for acquiring and managing the GPU clusters that underpin its model training and inference capacity.

Compute access is arguably the most consequential operational constraint OpenAI faces. The company has been navigating a complex web of data-center partnerships, including its ties to Microsoft (whose relationship has grown more complicated in recent months) and its broader Stargate infrastructure initiative. Placing a dedicated CTO over that function, rather than folding it under a broader engineering umbrella, suggests OpenAI is treating compute procurement and management as a strategic discipline in its own right.

Does Apple’s legal escalation change the picture for OpenAI?

Apple’s move to send preservation letters to dozens of former OpenAI employees represents a meaningful step up from simply filing a lawsuit. Preservation letters require recipients to retain documents and communications that could be relevant to litigation. Targeting ex-employees, rather than current ones, suggests Apple is trying to reconstruct what knowledge moved with people who left for OpenAI.

The backdrop, covered in prior editions, is Apple’s trade-secret suit filed in early July, which centers on allegations that OpenAI improperly used proprietary Apple technology and hired away key personnel. The letter campaign, reported by the Financial Times and aggregated via Daring Fireball, indicates Apple’s legal team is building a document record, which typically precedes discovery requests and depositions. This is a slow-burn legal risk, but it is becoming more operationally disruptive over time.

A $230 keypad as OpenAI’s hardware debut

The Codex Micro is a small RGB macropad built with hardware partner Work Louder. It lets users interact with AI coding agents via physical keys, reportedly including keycaps labeled “yeet” and “yolo.” At $230, it is a niche developer accessory rather than a consumer mass-market device.

The timing is notable. OpenAI released it July 15, five days after Apple’s lawsuit landed and one day after Bloomberg detailed the company’s separate Jony Ive-designed smart speaker project. Framing a tiny keypad as a hardware “launch” while a major consumer device is still in development creates an unusual optics situation, but the Codex Micro does mark OpenAI’s first physical product to market. For investors watching the company’s path toward potential public offering, the hardware ambitions remain a longer-dated story with significant execution risk still ahead.

Nothing in this update constitutes investment advice.