Morris II generative AI email worm (research PoC)
In controlled tests, self-replicating prompts spread between GenAI email assistants, demonstrating potential propagation and data-theft risks.
What happened
Adversarial self-replicating prompts make GenAI email assistants propagate malicious instructions to other agents (zero-click in lab), with potential spam/exfil side effects. Tested against GPT-4, Gemini Pro, LLaVA in controlled setups.
How it happened
Read the flow as text & sources
A catalog reconstruction of the supplied source account. Controlled research PoC against GenAI-powered email assistants; not observed in the wild.
- 01 · Research payload enters an email assistant
The controlled experiment places adversarial content in assistant-readable email.
Hunt: Untrusted content or a changed tool definition precedes a file read, command, or remote call outside the user's intended task.
Process logs cannot recover model intent. Remote connectors and in-process interpreters may produce no local child process. Prompt content alone is not a reliable signature.
Source for step 1 - 02 · Self-replicating instructions are processed
The model follows a prompt designed to reproduce itself.
Hunt: Untrusted content or a changed tool definition precedes a file read, command, or remote call outside the user's intended task.
Process logs cannot recover model intent. Remote connectors and in-process interpreters may produce no local child process. Prompt content alone is not a reliable signature.
Source for step 2 - 03 · Assistant propagates the payload
Generated email carries the adversarial instructions to another assistant.
Hunt: Related accounts or assistants repeat correlated posting, forwarding or tool actions beyond the expected user workflow.
Requires platform-side visibility. Similar content and synchronized activity can have legitimate explanations.
Source for step 3 - 04 · Lab demonstrates propagation
Potential spam and data-exfiltration effects are research outcomes, not confirmed real-world victims.
Hunt: Related accounts or assistants repeat correlated posting, forwarding or tool actions beyond the expected user workflow.
Requires platform-side visibility. Similar content and synchronized activity can have legitimate explanations.
Source for step 4
Why it matters
Reported impact: Potential spam and data-exfiltration effects are research outcomes, not confirmed real-world victims.
Check these source-linked conditions against your setup.
The controlled PoC propagates instructions between GenAI email assistants.
Conditions to check- The assistant must ingest attacker-controlled email content and have the demonstrated forwarding / retrieval capabilities.
- This is a controlled research proof of concept, not an observed in-the-wild worm campaign.
Source-linked applicability guidance. Confirm current scope in the original report.
What you can do
Protection & prevention
Treat retrieved email content as untrusted and constrain autonomous forwarding and data access.
Treat retrieved email/content as untrusted; isolation between agent mailboxes
Catalog defensive guidance based on the reported behavior. Confirm scope and current mitigations in the original source.
Detection queries
Draft query starters grouped by the attack stages where they apply.
Start with the logs you collect. Each hunt identifies its required visibility, field mapping and limits. Queries are catalog-authored drafts using reusable behavior profiles; test with known activity and your normal baseline before alerting.
01Research payload enters an email assistantUnexpected agent tool useCustom schema · Draft
The controlled experiment places adversarial content in assistant-readable email.
What to look for
Untrusted content or a changed tool definition precedes a file read, command, or remote call outside the user's intended task.
Visibility: Defender-side logs where collected
Required logs
- Agent tool-call audit records with session / tool / destination
- Tool-definition and approval changes
- Endpoint process, file-access and network evidence where available
Coverage limit. Process logs cannot recover model intent. Remote connectors and in-process interpreters may produce no local child process. Prompt content alone is not a reliable signature.
Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.
Sigma
title: Hunt - Unexpected agent tool use
id: c709e384-0e11-479e-ab4d-3a06e2431e4b
status: experimental
description: Untrusted content or a changed tool definition precedes a file
read, command, or remote call outside the user's intended task.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- agent.tool.call
- agent.tool.definition.change
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("agent.tool.call", "agent.tool.definition.change")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("agent.tool.call", "agent.tool.definition.change")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join input provenance, tool descriptions and calls by session ID and compare the action to the approved task.
- Inspect child processes and destinations; record whether execution was in-process, local, or through a remote SaaS connector.
02Self-replicating instructions are processedUnexpected agent tool useCustom schema · Draft
The model follows a prompt designed to reproduce itself.
What to look for
Untrusted content or a changed tool definition precedes a file read, command, or remote call outside the user's intended task.
Visibility: Defender-side logs where collected
Required logs
- Agent tool-call audit records with session / tool / destination
- Tool-definition and approval changes
- Endpoint process, file-access and network evidence where available
Coverage limit. Process logs cannot recover model intent. Remote connectors and in-process interpreters may produce no local child process. Prompt content alone is not a reliable signature.
Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.
Sigma
title: Hunt - Unexpected agent tool use
id: c709e384-0e11-479e-ab4d-3a06e2431e4b
status: experimental
description: Untrusted content or a changed tool definition precedes a file
read, command, or remote call outside the user's intended task.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- agent.tool.call
- agent.tool.definition.change
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("agent.tool.call", "agent.tool.definition.change")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("agent.tool.call", "agent.tool.definition.change")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join input provenance, tool descriptions and calls by session ID and compare the action to the approved task.
- Inspect child processes and destinations; record whether execution was in-process, local, or through a remote SaaS connector.
03Assistant propagates the payloadCoordinated messaging / propagationCustom schema · Draft
Generated email carries the adversarial instructions to another assistant.
What to look for
Related accounts or assistants repeat correlated posting, forwarding or tool actions beyond the expected user workflow.
Visibility: Platform / application records
Required logs
- Platform account and posting audit
- Agent session and message-provenance records
- Rate, recipient and content-similarity features
Coverage limit. Requires platform-side visibility. Similar content and synchronized activity can have legitimate explanations.
Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.
Sigma
title: Hunt - Coordinated messaging / propagation
id: c4001568-b68b-4887-a86f-6f05807640f6
status: experimental
description: Related accounts or assistants repeat correlated posting,
forwarding or tool actions beyond the expected user workflow.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- social.post
- message.forward
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("social.post", "message.forward")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("social.post", "message.forward")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Group actions by linked account, session, recipient and time rather than judging writing style.
- Compare propagation with legitimate campaigns and preserve whether the source describes a lab demonstration or observed abuse.
04Lab demonstrates propagationCoordinated messaging / propagationCustom schema · Draft
Potential spam and data-exfiltration effects are research outcomes, not confirmed real-world victims.
What to look for
Related accounts or assistants repeat correlated posting, forwarding or tool actions beyond the expected user workflow.
Use controlled-lab message and session records to investigate propagation; the research does not establish real-world victims.
Visibility: Platform / application records
Required logs
- Platform account and posting audit
- Agent session and message-provenance records
- Rate, recipient and content-similarity features
Coverage limit. Requires platform-side visibility. Similar content and synchronized activity can have legitimate explanations.
Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.
Sigma
title: Hunt - Coordinated messaging / propagation
id: c4001568-b68b-4887-a86f-6f05807640f6
status: experimental
description: Related accounts or assistants repeat correlated posting,
forwarding or tool actions beyond the expected user workflow.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
- Authorized activity matching this broad hunt; baseline and investigate
before alerting.
level: low
logsource:
category: application
product: openattacks_audit
definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
Target and CorrelationId from the required source logs first.
detection:
selection:
EventType:
- social.post
- message.forward
condition: selection
KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("social.post", "message.forward")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("social.post", "message.forward")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Group actions by linked account, session, recipient and time rather than judging writing style.
- Compare propagation with legitimate campaigns and preserve whether the source describes a lab demonstration or observed abuse.
Technical detailsTooling, impact & MITRE ATLAS
Attack context
- AI's role
- Agent / MCP as target
Tooling as reported
AI products, models and artifacts
- ChatGPT-4
- Gemini Pro
- LLaVA
- RAG email assistants
Detection rules
- Review attack-specific indicatorsGUIDANCE
Monitor for the techniques described in Morris II generative AI email worm (research PoC). Refer to sources for specific IOCs and TTPs.
Source defensive notes
Treat retrieved email/content as untrusted; isolation between agent mailboxes
MITRE ATLAS
AI-specific catalog annotations based on the reported behavior; these are not MITRE case-study assignments.
Email content injects instructions into assistants in a controlled experiment. Source
The prompt is designed to replicate into assistant output. Source
Sources & research credit
2 sourcesResearch by Cohen, Bitton, Nassi et al. · Academic / Cornell Tech et al.
Record & downloads
- Source record name
- Morris II generative AI email worm (research PoC)
- Entry ID
- morris-ii-ai-worm-poc-2024-03
- Catalog review
- October 20, 2026
- Category
- attacks-on-ai
- AI involvement confidence
- unverified