Research demo
Computer-use TOCTOU: swapping UI elements during inference so the agent clicks something else
AIR-2026-0141 AIR record · 2026-06-anthropic-computer-use-toctou
From the supporting source
The Agent will think it clicks the Continue button, but in reality it clicks "Send" to send an arbitrary email.embracethered.com
The record
Disclosed
2026-06-25
Vendor
Anthropic
Product
Claude Computer Use
Risk domain
Agent behaviour › Unsafe actuation
Also touches
Data exfiltration channels
System type
Generative model
Who acted
Agent acted — the agent clicked the swapped UI element and sent the email
Attack surface
Computer-use agent
Autonomy level
Semi-autonomous
Initial vector
Malicious web ui
Impact type
Unauthorized action, Data exfiltration
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
computer-use agents that act on stale screenshots
Capability abused
Defense & remediation
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
Anthropic added pixel-change verification before taking actions (per its Cowork announcement)
Taxonomy & references
Curator notes
Johann Rehberger. Class of bug applies to any screenshot-then-act computer-use loop, not only Anthropic's.
Supporting sources
Cite as
Enkrypt AI. "Computer-use TOCTOU: swapping UI elements during inference so the agent clicks something else". AIR, AIR-2026-0141. https://agentincident.vercel.app/i/AIR-2026-0141/ (accessed 2026-10-19).