Email prompt injection against mail AI assistants (Sublime, Barracuda)
Sublime Security (Harley Halsey, Brandon Murphy, Randy Pargman) primary research documents invoice fraud campaign hiding forward this email with its PDF to ap.confirm+<controller>@<target> instructions in…
What happened
Sublime Security (Harley Halsey, Brandon Murphy, Randy Pargman) primary research documents invoice fraud campaign hiding forward this email with its PDF to ap.confirm+<controller>@<target> instructions in zero-size HTML span elements and Unicode Tag characters invisible to humans. Barracuda (Merium Khalid, Nov 7) via Cybersecurity Insiders reports recruiting email hiding HTML playbook telling assistant to forward to all-staff mailing lists for self-propagation and search mailbox for seed phrases, private keys, wallet data to exfiltrate to outside Gmail accounts. Both campaigns target email AI assistants reading/processing messages on behalf of users. Unattributed actors; confirmed in wild by Sublime and Barracuda independently.
Email AI assistants manipulated via hidden HTML prompt injection to forward emails, search mailboxes for sensitive data, and self-propagate across organization
How it happened
Read the flow as text & sources
A catalog reconstruction of the supplied source account. Unknown steps remain explicit; an edge describes the reported relationship between steps.
- 01 · Craft hidden prompt injection
Attacker embeds instructions in zero-size HTML spans and Unicode Tag characters invisible to users
Hunt: An impersonating message, ad or website leads to a suspicious download, consent request, pasted command or payment instruction.
Polished text, AI branding and a novel domain do not prove maliciousness or AI generation. Follow-on evidence matters.
Source for step 1 - 02 · Deliver phishing email
Invoice fraud or recruiting email delivered to target organization
Hunt: An impersonating message, ad or website leads to a suspicious download, consent request, pasted command or payment instruction.
Polished text, AI branding and a novel domain do not prove maliciousness or AI generation. Follow-on evidence matters.
Source for step 2 - 03 · Email AI assistant processes message
AI assistant reads email including hidden instructions
Hunt: An agent ingests external content or an untrusted session identifier before a tool action or configuration change outside the approved task.
Requires application instrumentation and a custom field mapping. Input exposure does not prove that the model followed instructions, executed a tool or bypassed approval. Avoid collecting sensitive prompt bodies unless authorized.
Source for step 3 - 04 · AI forwards email or searches mailbox
Assistant follows hidden instructions to forward email with PDF to attacker-controlled address or search for sensitive data
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 4 - 05 · Self-propagate via all-staff forwards
Barracuda case: AI forwards to all-staff lists for organizational spread
Hunt: Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
Source for step 5
Why it matters
Reported entry point: Phishing emails with hidden HTML prompt injection targeting email AI assistants. Reported impact: Barracuda case: AI forwards to all-staff lists for organizational spread; reported targets include Organizations using email AI assistants; finance teams (invoice fraud); general workforce (recruiting fraud); applicability to your setup has not been assessed, so verify affected versions, permissions, and configuration in the original report.
What you can do
Protection & prevention
Strip hidden HTML elements from emails before AI processing; sanitize Unicode Tag characters; implement AI assistant output validation; restrict AI assistant forwarding capabilities; require human approval for bulk forwards; monitor for unusual mailbox search patterns; educate users on prompt injection risks; implement strict email content security policies
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.
01Craft hidden prompt injectionLure → user action / deliveryCustom schema · Draft
Attacker embeds instructions in zero-size HTML spans and Unicode Tag characters invisible to users
What to look for
An impersonating message, ad or website leads to a suspicious download, consent request, pasted command or payment instruction.
Visibility: Defender-side logs where collected
Required logs
- Email / secure-web-gateway records
- Browser redirects, downloads and URL reputation
- User reports and downstream execution / identity events
Coverage limit. Polished text, AI branding and a novel domain do not prove maliciousness or AI generation. Follow-on evidence matters.
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 - Lure → user action / delivery
id: e0056b49-9905-485b-ac12-2aa076a7bc82
status: experimental
description: An impersonating message, ad or website leads to a suspicious
download, consent request, pasted command or payment instruction.
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:
- message.link.click
- web.download
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 ("message.link.click", "web.download")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("message.link.click", "web.download")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Reconstruct redirect and delivery provenance using original report indicators, without visiting payload URLs.
- Correlate the click with the actual download, consent or execution; verify the sender through an independent channel.
02Deliver phishing emailLure → user action / deliveryCustom schema · Draft
Invoice fraud or recruiting email delivered to target organization
What to look for
An impersonating message, ad or website leads to a suspicious download, consent request, pasted command or payment instruction.
Visibility: Defender-side logs where collected
Required logs
- Email / secure-web-gateway records
- Browser redirects, downloads and URL reputation
- User reports and downstream execution / identity events
Coverage limit. Polished text, AI branding and a novel domain do not prove maliciousness or AI generation. Follow-on evidence matters.
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 - Lure → user action / delivery
id: e0056b49-9905-485b-ac12-2aa076a7bc82
status: experimental
description: An impersonating message, ad or website leads to a suspicious
download, consent request, pasted command or payment instruction.
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:
- message.link.click
- web.download
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 ("message.link.click", "web.download")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("message.link.click", "web.download")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Reconstruct redirect and delivery provenance using original report indicators, without visiting payload URLs.
- Correlate the click with the actual download, consent or execution; verify the sender through an independent channel.
03Email AI assistant processes messageUntrusted content enters an agent sessionCustom schema · Draft
AI assistant reads email including hidden instructions
What to look for
An agent ingests external content or an untrusted session identifier before a tool action or configuration change outside the approved task.
Visibility: Application / agent audit records
Required logs
- Input provenance, source type and request/session identifier
- Captured tool responses or retrieved content where collection is approved
- Subsequent tool calls, configuration changes and approval decisions
Coverage limit. Requires application instrumentation and a custom field mapping. Input exposure does not prove that the model followed instructions, executed a tool or bypassed approval. Avoid collecting sensitive prompt bodies unless authorized.
Custom-schema candidate for input exposure only. Map agent.input.ingest and agent.session.resume to your application request and session logs. CorrelationId must link subsequent actions. EventType values are proposed normalization names, not native vendor events.
Sigma
title: Hunt - Untrusted content enters an agent session
id: 2d89d7ae-9d44-4dc6-a95b-f9c8de7b8352
status: experimental
description: An agent ingests external content or an untrusted session
identifier before a tool action or configuration change outside the approved
task.
author: MagicSword OpenAttacks catalog
date: 2026-11-04
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.input.ingest
- agent.session.resume
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.input.ingest", "agent.session.resume")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("agent.input.ingest", "agent.session.resume")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join the ingested input to later actions by session or request ID; preserve the original content as evidence.
- Compare the resulting action with the approved task and baseline. Receiving suspicious text alone is not successful injection.
04AI forwards email or searches mailboxUnexpected agent tool useCustom schema · Draft
Assistant follows hidden instructions to forward email with PDF to attacker-controlled address or search for sensitive data
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.
05Self-propagate via all-staff forwardsSensitive data → unusual destinationCustom schema · Draft
Barracuda case: AI forwards to all-staff lists for organizational spread
What to look for
Sensitive reads are followed by outbound transfer, a public artifact, or a tool call to a destination outside the approved workflow.
Visibility: Defender-side logs where collected
Required logs
- Sensitive resource access / DLP records
- Proxy, egress and cloud upload events
- Agent / SaaS audit with correlation IDs
Coverage limit. An outbound connection or high byte count alone is not exfiltration. Encrypted traffic may hide content, and bulk legitimate exports are common.
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 - Sensitive data → unusual destination
id: 04b60968-3179-4b1f-ae13-6f57c2446077
status: experimental
description: Sensitive reads are followed by outbound transfer, a public
artifact, or a tool call to a destination outside the approved 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:
- data.export
- agent.external.send
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 ("data.export", "agent.external.send")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated descSPL
sourcetype="openattacks:audit"
EventType IN ("data.export", "agent.external.send")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Join the resource read and outbound action by session, identity or process, preserving a bounded time window.
- Inspect destination ownership, payload classification and the user's intended action; baseline sanctioned sharing.
Technical detailsTooling, impact & MITRE ATLAS
Attack context
- AI's role
- Agent / MCP as target
- Initial access
- Phishing emails with hidden HTML prompt injection targeting email AI assistants
Tooling as reported
AI products, models and artifacts
- Email AI assistants (unspecified vendors)
- Mailbox automation agents
Other tooling
- Zero-size HTML span elements
- Unicode Tag characters
- HTML playbook instructions
Targets and reported impact
- Targets / sectors
- Organizations using email AI assistants; finance teams (invoice fraud); general workforce (recruiting fraud)
- Reported scale / exposure
- Active Nov 2026; confirmed by two independent security vendors
- Target geography
- Multiple regions
Reported attribution
Unattributed campaigns; confirmed active in wild Nov 2026 by Sublime and Barracuda
Structured attribution
- Actor
- Unattributed campaigns
- Country
- Unknown
- Motivation
- financially-motivated
- Attributed by
- Sublime Security / Barracuda
- Attribution confidence
- low
Detection rules
- Email AI assistant security controlsGUIDANCE
Strip zero-size HTML elements before AI processing; sanitize Unicode Tag characters (U+E0000–U+E007F); detect hidden text with display:none, font-size:0, opacity:0; validate AI assistant actions (forwards, searches); require human approval for bulk operations; monitor for ap.confirm+ email patterns; alert on mailbox searches for seed phrase, private key, wallet; log all AI assistant forwarding actions; implement AI output validation
Activity timeline
- First seen
- 2026-11-01
- Last updated
- 2026-11-07
Updates
- 2026-11-06 — Sublime Security
Sublime reports invoice fraud campaign using zero-size HTML spans and Unicode Tag characters to inject hidden instructions targeting email AI assistants
- 2026-11-07 — Barracuda / Cybersecurity Insiders
Barracuda documents recruiting email hiding HTML playbook for all-staff self-propagation and mailbox search for cryptocurrency data
Source defensive notes
Strip hidden HTML elements from emails before AI processing; sanitize Unicode Tag characters; implement AI assistant output validation; restrict AI assistant forwarding capabilities; require human approval for bulk forwards; monitor for unusual mailbox search patterns; educate users on prompt injection risks; implement strict email content security policies
MITRE ATLAS
AI-specific catalog annotations based on the reported behavior; these are not MITRE case-study assignments.
Hidden HTML instructions inject prompts into email AI assistant processing Source
Email AI assistants manipulated to forward emails and search mailboxes without user awareness Source
AI assistants instructed to exfiltrate seed phrases, private keys, and wallet data Source
Sources & research credit
2 sourcesResearch by Harley Halsey, Brandon Murphy, Randy Pargman (Sublime); Merium Khalid (Barracuda) · Sublime Security, Barracuda Networks
- Sublime: Invoice Fraud Campaign Adds Prompt Injection Aimed at Email AI Agents sublime.security
- Phishing Has a New Target - The AI Assistant Reading Your Email www.cybersecurity-insiders.com
Sublime Security blog (primary source) documents invoice fraud campaign with zero-size span and Unicode Tag injection; Barracuda report via Cybersecurity Insiders adds recruiting fraud with all-staff self-propagation and mailbox search for crypto data; both confirm active attacks Nov 2026
Record & downloads
- Source record name
- Email prompt injection against mail AI assistants (Sublime, Barracuda)
- Entry ID
- email-prompt-injection-ai-assistants-2026-11
- Catalog review
- November 10, 2026
- Category
- attacks-on-ai
- AI involvement confidence
- confirmed