Research demo
Anthropic ships Claude for Chrome with a measured 23.6% prompt-injection attack success rate pre-mitigation
AIR-2025-0114 AIR record · 2025-08-anthropic-claude-for-chrome-injection-evals
From the supporting source
Prompt injection attacks can cause AIs to delete files, steal data, or make financial transactions.claude.com
The record
Disclosed
2025-08-25
Vendor
Anthropic
Product
Claude for Chrome
Risk domain
Agent behaviour › Unsafe actuation
Also touches
Data exfiltration channels, Confused deputy
System type
Generative model
Who acted
Agent acted — the browser agent deleted email and submitted forms under injection
Attack surface
Browser agent
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection web and email
Impact type
Data destruction, Data exfiltration, Financial loss
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
initial 1,000-user research preview
Capability abused
Defense & remediation
Recorded defense
site-level permissions, action confirmations for publish/purchase/personal-data actions, blocked high-risk site categories, classifiers
Guardrail kinds
Input classifier, Human approval, Permission scope
Guardrail outcome
Silent
Fix shipped
mitigations reduced autonomous-mode ASR from 23.6% to 11.2%; browser-specific attack challenge set from 35.7% to 0%
Taxonomy & references
Curator notes
Vendor self-disclosure rather than an external finding, but included because it is the clearest published measurement of residual browser-agent injection risk in a shipped product: 11.2% of attacks still succeed after mitigations, across 123 test cases / 29 attack scenarios. Worked example: an email impersonating the employer's security team told Claude to delete the user's mailbox 'for mailbox hygiene' with 'no additional confirmation required' - undefended Claude complied.
Supporting sources
Cite as
Enkrypt AI. "Anthropic ships Claude for Chrome with a measured 23.6% prompt-injection attack success rate pre-mitigation". AIR, AIR-2025-0114. https://agentincident.vercel.app/i/AIR-2025-0114/ (accessed 2026-10-19).