Disclosed vulnerability
Google Jules is vulnerable to invisible Unicode Tag prompt injection
AIR-2025-0099 AIR record · 2025-08-google-jules-invisible-injection
From the supporting source
The latest Gemini models quite reliably interpret hidden Unicode Tag characters as instructions.embracethered.com
The record
Disclosed
2025-08-15
Vendor
Google
Product
Google Jules
Risk domain
Access control and permissions › Privilege escalation
Also touches
Supply chain vulnerabilities
System type
Generative model
Who acted
Agent acted — Jules acted on invisible Unicode instructions and wrote a backdoor
Attack surface
Coding agent
Autonomy level
Fully autonomous
Initial vector
Invisible unicode injection
Impact type
Remote code execution, Backdoor insertion
Confidence
High
Submitted by
EAI Internal
Crosswalk
- OWASP Agentic Top 10 (2026)
- ASI03 – Identity and Privilege Abuse
- MITRE ATLAS
- AML.T0055 – Unsecured Credentials
- NIST AI RMF
- GOVERN 6.1
- EU AI Act
- Article 13
- ISO/IEC
- TR 24028; 42001; 23894
- AIUC-1
- B. Security
- Scope
- tool misuse
- Mitigation
- Apply least-privilege principle; Regular policy audits.
- Taxonomy
- Enkrypt AI agent risk taxonomy · Black-Box Red Teaming of Agentic AI: A Taxonomy-Driven Framework for Automated Risk Discovery
Blast radius
Jules users processing GitHub issues and untrusted data
Capability abused
Defense & remediation
Recorded defense
plan review by the user
Guardrail kinds
Unknown
Guardrail outcome
Silent
Fix shipped
not fixed at publication; first reported to Google 2024-02-22, accepted but unmitigated, and Jules went GA without remediation
Taxonomy & references
Curator notes
Month of AI Bugs 2025.
Supporting sources
Cite as
Enkrypt AI. "Google Jules is vulnerable to invisible Unicode Tag prompt injection". AIR, AIR-2025-0099. https://agentincident.vercel.app/i/AIR-2025-0099/ (accessed 2026-10-19).