Agent Incident RegistryAIR
v1 · updated 19 Oct 2026
Research demo

AgentHopper: a universal prompt-injection payload that hops between shipped coding agents

AIR-2025-0128 AIR record · 2025-08-agenthopper-ai-virus
From the supporting source
The idea was to have one prompt injection payload that would operate across agents and exploit them accordingly.embracethered.com
The record
Disclosed
2025-08-29
Vendor
GitHub/Microsoft; Amazon/AWS; Sourcegraph
Product
GitHub Copilot; Amazon Q Developer; AWS Kiro; Amp Code
Risk domain
Access control and permissions › Privilege escalation
Also touches
Confused deputy
System type
Generative model
Who acted
Agent acted — the payload made each agent write config and execute, hopping onward
Attack surface
Coding agent
Autonomy level
Semi-autonomous
Initial vector
Indirect prompt injection repo content
Impact type
Code execution, Self propagation, Persistence, 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 and any Git project the infected agent touches
Capability abused
config_file_writeauto_approve_togglecode_executiongit_operationsfile_writeconfig_modificationgit_push
Defense & remediation
Recorded defense
per-agent approval gates and configuration protections
Guardrail kinds
Human approval
Guardrail outcome
Silent
Fix shipped
all targeted vulnerabilities patched by the vendors
Curator notes
Johann Rehberger, 29 Aug 2025. One payload fingerprints which coding agent is reading it and fires the matching exploit (e.g. writing chat.tools.autoApprove into .vscode/settings.json for Copilot, a fake MCP server config for Amp Code). Research PoC; the article does not indicate the code was released. This is the clearest cross-vendor agent-virus demonstration against shipped products prior to the 2026 OpenClaw worm wave. | Merged from duplicate id `2025-08-embracethered-agenthopper`. Also reported as: 'AgentHopper: a proof-of-concept AI virus that self-propagates across coding agents via prompt injection'.
Cite as
Enkrypt AI. "AgentHopper: a universal prompt-injection payload that hops between shipped coding agents". AIR, AIR-2025-0128. https://agentincident.vercel.app/i/AIR-2025-0128/ (accessed 2026-10-19).