Safety failure
Realized harm
Cursor agent ran destructive commands while Plan Mode was active, deleting ~70 git-tracked files and ignoring an explicit stop instruction
AIR-2025-0191 AIR record · 2025-12-cursor-plan-mode-constraint-violation
A real party was affected. This incident caused actual consequences, not a demonstrated capability.
From the supporting source
get everything into the correct state to run and DO NOT RUN ANYTHINGforum.cursor.com
The record
Disclosed
2025-12-08
Occurred
2025-12-08
Vendor
Anysphere / Cursor
Product
Cursor (Plan Mode, Claude Opus)
Risk domain
Agent behaviour › Unsafe actuation
Also touches
Uncontrolled resource consumption
System type
Generative model
Who acted
Agent acted — the agent ran destructive commands and deleted ~70 tracked files
Attack surface
Coding agent
Autonomy level
Copilot with approval
Initial vector
Autonomous agent action
Impact type
Data destruction, Service disruption
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
two developer machines; ~70 tracked files deleted, test processes killed on both hosts
Capability abused
Defense & remediation
Recorded defense
Cursor Plan Mode read-only constraint ('You MUST NOT make any edits or run any non-readonly tools until explicitly instructed') plus a user instruction in caps
Guardrail kinds
Prompt instruction, Permission scope
Guardrail outcome
Silent
Fix shipped
Cursor staff acknowledged a critical bug in Plan Mode constraint enforcement and escalated it
Taxonomy & references
Curator notes
Best-documented case of an agent circumventing its own approval gate in normal operation: the gate is implemented as a system-prompt instruction, so it is advisory rather than enforced. The agent also ignored an explicit machine-scoping constraint, acting on Machine B after being told to touch only Machine A. Cursor staff member 'Dean Rie' called it 'a violation of Plan Mode's read-only constraint'. Multiple sibling forum threads report Plan Mode silently switching to Agent Mode.
Cite as
Enkrypt AI. "Cursor agent ran destructive commands while Plan Mode was active, deleting ~70 git-tracked files and ignoring an explicit stop instruction". AIR, AIR-2025-0191. https://agentincident.vercel.app/i/AIR-2025-0191/ (accessed 2026-10-19).