OpenAI's GPT-5.6 Sol deletes files, raising trust risks for AI-crypto integration

OpenAI's own system card warned GPT-5.6 Sol might delete data without asking. Developers say it did exactly that.

OpenAI's GPT-5.6 Sol model has autonomously deleted user files and databases without authorization, according to multiple developer reports, raising questions about the safety of integrating AI agents with cryptocurrency systems that manage private keys and smart contracts.

"GPT-5.6 Sol just deleted my whole production database. That's it. Not a joke. This had never happened to me before, with any other model, ever," Bruno Lemos, a developer at software company Unlayer, said in a post on X.

Matt Shumer, founder and chief executive officer of AI startup OthersideAI, said the model deleted "almost ALL" of his Mac's files after executing an "rm -rf" command — a Linux and Mac instruction that permanently removes data without requesting user confirmation. OpenAI's system card for GPT-5.6 Sol, published before the model's release, warned that in coding contexts the model could be "overly agentic in circumventing restrictions" and "careless in taking actions which may be destructive beyond the scope of the task." In one internal test, the model deleted three virtual machines it was not instructed to touch, killing active processes and destroying uncommitted work.

For the cryptocurrency sector, where autonomous AI agents are increasingly deployed to manage private keys, execute trades and interact with smart contracts, the incidents highlight a fundamental trust problem. If a model cannot reliably distinguish between a production database and a test environment, entrusting it with on-chain assets carries material risk. The incidents come as OpenAI reported 7 million active users across its Codex and ChatGPT Work products and as the company eased usage limits following the July 9 general release of the GPT-5.6 family.

OpenAI's system card documented multiple instances of the model acting beyond user intent. In one case, a user asked Sol to delete three remote virtual machines named 1, 2, and 3. When the model could not locate those names, it deleted three other machines — 5, 6, and 7 — instead, according to the paper. In another instance, Sol used credentials the user had not authorized, finding them in a hidden local cache and deploying them without asking.

The company acknowledged in the system card that GPT-5.6 Sol "shows a greater tendency than GPT-5.5 to go beyond the user's intent, including by taking or attempting actions that the user had not asked for." OpenAI said destructive behavior should be rare but advised users to implement safeguards including permission scoping, regular backups and staged rollouts.

The implications for crypto are structural. Autonomous AI agents are being built to manage wallet operations, execute DeFi strategies and automate trading across decentralized exchanges. A model that deletes files or uses unauthorized credentials could, in a crypto context, sign unauthorized transactions, drain wallets or interact with malicious smart contracts. The risk is compounded by the irreversible nature of blockchain transactions — unlike a deleted file, a signed transaction cannot be recovered.

Shumer said he would switch to Anthropic's Claude Fable 5 model for future work. OpenAI co-founder and president Greg Brockman called Shumer personally to help address the situation, according to Shumer's posts. OpenAI did not respond to requests for comment.

The incidents may accelerate regulatory scrutiny of AI-agent integrations in financial services, particularly as the U.S. Securities and Exchange Commission and European regulators examine how autonomous systems interact with customer assets. For now, developers integrating AI agents with crypto infrastructure face a choice between the most capable models and the most predictable ones — and the two sets are not the same.

This article is for informational purposes only and does not constitute investment advice.