This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A user named Tassos M reported that Google's 'Antigravity' agentic development platform, while running in 'Turbo mode', incorrectly executed a command that wiped his entire D: drive partition. The AI admitted to the error, which occurred due to a path parsing failure during a cache cleanup task. The user was unable to recover the deleted files, highlighting significant risks associated with granting AI agents unrestricted terminal access.
A user identified publicly as Tassos M reported that Google's Antigravity IDE, while attempting to clear a project cache, executed rmdir /s /q d:\ and deleted an entire D: drive partition instead of the intended folder. The user said recovery largely failed, and Google later said it was aware of the report and investigating.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
The user ran the AI in 'Turbo mode', granting it unrestricted terminal access to execute commands without manual confirmation, representing unsafe use of an uncontained agent.
The incident was primarily caused by a technical failure where the AI system suffered a path parsing error and executed an incorrect, destructive command, demonstrating a lack of robustness in critical operations.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The incident was caused by a path parsing error and subsequent destructive command execution by the deployed Antigravity AI system, which was an unintended outcome of its cache cleanup goal.
Google's Antigravity agentic platform running in 'Turbo mode' catastrophically deleted a user's hard drive partition due to a path parsing error. While highlighting significant risks of granting AI agents unrestricted terminal access, the incident is a consumer safety and software robustness failure with negligible direct implications for national security.
The platform is described as an 'agentic development platform' and an 'AI agent' that can execute terminal commands automatically in Turbo mode to achieve user-defined goals, such as clearing a project cache.
Limited Risk: The system is a Gemini 3-based interactive AI agent, which falls under the category of chatbots and interactive AI systems requiring transparency under the EU AI Act.
The AI system, Antigravity, is explicitly described as an agentic development platform used to help users build applications and write code.
The reports provide consistent, detailed accounts of the incident, including direct quotes from the user, the AI's console logs, and Google's acknowledgment of the investigation. The technical mechanism of the failure (path parsing error leading to rmdir /s /q d:\) is clearly explained.
None that are plausible; the evidence strongly points to an AI command execution failure.
Technology
Path Parsing Error
AI incorrectly resolved the project cache path to the root D drive.
Aggressive Command Generation
AI generated 'rmdir /s /q' command which deletes recursively and quietly.
Lack of Sandbox Isolation
The platform ran commands directly on the host OS instead of a container.
Data Inputs
Faulty Path Data Parsing
AI failed to properly parse the target directory string for cache cleanup.
Human Factors
Use of Turbo Mode
User enabled Turbo mode, allowing automatic execution without manual approval.
Lack of Local Isolation
User ran the powerful agent directly on host machine with critical data.
Overreliance on AI Tool
Non-developer user trusted the AI to manage system files safely.
Process and Methods
No Command Verification
No verification step existed for potentially destructive system commands.
Direct Host Execution
Execution method allowed direct terminal access without virtualization.
Management
Marketing to Non-Developers
Google marketed agentic tools to hobbyists without adequate safety warnings.
Inadequate Safety Guardrails
Management released the platform without restricting destructive commands.
Regulatory Environment
Lack of AI Safety Standards
No industry standards mandate sandboxing for agentic developer tools.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
