A Reported GPT-5.6
Summary
Former HyperWrite CEO Matt Shumer reported on X that while testing OpenAI's GPT-5.6 Sol in "Ultra" mode, an AI agent allegedly deleted nearly all files on his Mac. The incident, described as a "freak accident," involved a `review subagent` executing the command `rm -rf /Users/mattsdevbox` after incorrectly expanding `\$`. This event highlights a critical shift in AI safety: when models transition from merely providing advice to directly executing actions, errors escalate from incorrect answers to destructive system commands with real-world consequences. The reported local data loss underscores the potential for agents with broad filesystem access to cause significant damage, extending beyond local files to potentially impact databases, credential stores, or production pipelines.
Key takeaway
For AI Security Engineers or MLOps Engineers deploying autonomous agents, this incident underscores that your security posture must shift from "trust the model" to "control the action." You must implement least privilege, ensuring agents default to workspace-only access and require human approval for all destructive commands. Isolate high-privilege tasks in containers or VMs, and block dangerous shell patterns like `rm -rf` at the execution boundary to prevent catastrophic data loss.
Key insights
AI agents with execution capabilities transform safety failures from wrong answers into destructive system commands.
Principles
- Excessive agency increases incident severity from low-probability mistakes.
- Agents lack human implicit constraints, leading to task scope drift.
- Errors can propagate and normalize across multi-agent chains.
In practice
- Default agent filesystem access to project workspace only.
- Implement human approval for all destructive file operations.
- Block dangerous command patterns like `rm -rf` at runtime.
Topics
- AI Agents
- Agent Security
- Runtime Security
- Least Privilege
- Data Loss Prevention
- Execution Control
Best for: AI Security Engineer, AI Engineer, MLOps Engineer
Related on AIssential
See Counsel's argued verdicts on the open AI decisions leaders are weighing →
Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence on Medium.