Disclosed vulnerability
Cross-agent privilege escalation: one hijacked coding agent rewrites another agent's config to free it
AIR-2025-0136 AIR record · 2025-09-cross-agent-privilege-escalation
From the supporting source
What starts as a single indirect prompt injection can quickly escalate into a multi-agent compromise, where one agent 'frees' another agent and sets up a loop of escalating privilege and control.embracethered.com
The record
Disclosed
2025-09-24
Vendor
multiple
Product
GitHub Copilot, Claude Code, AWS Kiro (cross-agent chain)
Risk domain
Access control and permissions › Privilege escalation
Also touches
Confused deputy
System type
Generative model
Who acted
Agent acted — one agent rewrote another agent's config to free it, then executed
Attack surface
Multi-agent system
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection source code
Impact type
Privilege escalation, Remote code execution
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
developer machines running more than one coding agent over the same workspace
Capability abused
Defense & remediation
Recorded defense
per-agent approval prompts
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
not fixed as a class; MSRC considered the demo not severe enough to require immediate security servicing
Taxonomy & references
Curator notes
Johann Rehberger. Key insight: agents write each other's config files (.vscode/settings.json, .mcp.json, CLAUDE.md) without approval.
Supporting sources
Cite as
Enkrypt AI. "Cross-agent privilege escalation: one hijacked coding agent rewrites another agent's config to free it". AIR, AIR-2025-0136. https://agentincident.vercel.app/i/AIR-2025-0136/ (accessed 2026-10-19).