Disclosed vulnerability
Brave: hidden white-on-white instructions hijack Mozilla Tabstack's cloud browser agent to exfiltrate conversation history
AIR-2026-0131 AIR record · 2026-06-mozilla-tabstack-injection
From the supporting source
The agent never summarized the page. Instead, it executed all three injected steps in sequence.brave.com
The record
Disclosed
2026-06-08
Vendor
Mozilla; Cotypist
Product
Mozilla Tabstack (cloud-hosted browser agent); Cotypist (on-device macOS autocomplete)
Risk domain
Privacy › Data exfiltration channels
Also touches
Credential theft, Confused deputy
System type
Generative model
Who acted
Agent acted — the cloud browser agent navigated and submitted the conversation history
Attack surface
Browser agent
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection web page
Impact type
Data exfiltration, Credential exposure
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI02 – Tool Misuse and Exploitation
- MITRE ATLAS
- AML.T0024 – Exfiltration via AI Inference API
- NIST AI RMF
- MAP 4.2
- EU AI Act
- Article 10
- ISO/IEC
- TR 24028; 23894
- AIUC-1
- A. Data & Privacy
- Scope
- tool misuse
- Mitigation
- Monitor network traffic for anomalous patterns; Restrict egress destinations.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Tabstack users summarizing attacker-controlled pages; Cotypist users opening attacker-supplied local documents
Capability abused
Defense & remediation
Recorded defense
none effective
Guardrail kinds
None
Guardrail outcome
Not applicable
Fix shipped
unknown
Taxonomy & references
Curator notes
Two 2026 cases in one Brave writeup, showing the class is not confined to Comet/Atlas. (1) Mozilla Tabstack: asked to summarize a page, the cloud agent instead followed white-on-white instructions, navigated to an attacker domain, pulled its own conversation history and POSTed it via a form. (2) Cotypist: a fully on-device macOS autocomplete model was steered by instructions inside local documents into suggesting false content and potentially surfacing credentials inline - Brave notes small local models are at higher risk of failing to separate instructions from data.
Supporting sources
Cite as
Enkrypt AI. "Brave: hidden white-on-white instructions hijack Mozilla Tabstack's cloud browser agent to exfiltrate conversation history". AIR, AIR-2026-0131. https://agentincident.vercel.app/i/AIR-2026-0131/ (accessed 2026-10-19).