Research demo
PajaMAS: hijacking multi-agent systems built on Google's Agent Development Kit
AIR-2025-0070 AIR record · 2025-07-trailofbits-mas-hijacking-adk
From the supporting source
This is the fault in our agents: they're confused deputies laundering malicious data from other agents.blog.trailofbits.com
The record
Disclosed
2025-07-31
Vendor
Google
Product
Google Agent Development Kit (ADK)
Risk domain
Access control and permissions › Privilege escalation
Also touches
Confused deputy
System type
Generative model
Who acted
Agent acted — injected inter-agent messages drove code execution in the MAS
Attack surface
Multi-agent system
Autonomy level
Semi-autonomous
Initial vector
Inter agent message injection
Impact type
Arbitrary code execution, Agent hijack
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
Multi-agent systems using ADK topologies
Capability abused
Taxonomy & references
Curator notes
Author Suha Sabi Hussain. Released seven demo systems (simple_mas, tool_mas, agent_memory_mas, trifecta_mas, cycle_mas, url_anchor_mas, defense_mas).
Supporting sources
Cite as
Enkrypt AI. "PajaMAS: hijacking multi-agent systems built on Google's Agent Development Kit". AIR, AIR-2025-0070. https://agentincident.vercel.app/i/AIR-2025-0070/ (accessed 2026-10-19).