Safety failure
Gemini CLI invented a file-deletion task in response to a plain `ls` command
AIR-2025-0185 AIR record · 2025-12-google-gemini-cli-hallucinated-deletion-task
From the supporting source
The original task was `ls`, gemini hallucinated a task to delete files.github.com
The record
Disclosed
2025-12-02
Occurred
2025-12-02
Vendor
Google
Product
Gemini CLI
Risk domain
Agent behaviour › Unsafe actuation
System type
Generative model
Who acted
Agent acted (contested) — the agent invented and prepared a deletion task on its own
Attack surface
Coding agent
Autonomy level
Semi-autonomous
Initial vector
Autonomous agent action
Impact type
Near miss data destruction
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI01 – Agent Goal Hijack
- MITRE ATLAS
- AML.T0048 – External Harms
- NIST AI RMF
- MEASURE 2.6; MANAGE 1.3
- EU AI Act
- Annex III §1(c)
- ISO/IEC
- TR 24028; 24029-1; 23894
- AIUC-1
- C. Safety
- Scope
- agent misuse
- Mitigation
- Require confirmation for destructive actions; Implement dry-run modes; Use layered safety controls.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
single user session; user caught it before deletion completed
Capability abused
Defense & remediation
Recorded defense
user review of proposed commands
Guardrail kinds
Unknown
Guardrail outcome
Silent
Taxonomy & references
Curator notes
impact_realized=false: the agent announced 'I need to delete all uz_ prefixed files in /home/user/backup' and prepared deletion commands after a read-only request, but the reporter does not establish that files were actually lost. Useful as evidence that goal invention, not just faulty execution, is a live failure mode. A separate, related report (issue #16595, 'Unsolicited File Operations After list files') describes the same pattern.
Supporting sources
Cite as
Enkrypt AI. "Gemini CLI invented a file-deletion task in response to a plain `ls` command". AIR, AIR-2025-0185. https://agentincident.vercel.app/i/AIR-2025-0185/ (accessed 2026-10-19).