Attack scenario · AI security

Malicious content influencing an agent’s privileged tool invocation

How can a web page or ticket make an AI agent take a harmful action? An agent browsing tickets and web pages encounters crafted content that steers it into calling a privileged tool — such as sending data externally or deleting records.

Fictional, educational scenarioUnauthorized actionsData exposureService disruptionLast updated · ATT&CK v20.1

Attack path · 4 stages

  1. 1

    Submit crafted content

    T1698AML.T0051.001

  2. 2

    Agent queries customer data

    T1213AML.T0053

  3. 3

    Agent sends data externally

    T1700T1567AML.T0086

  4. 4

    Agent closes or deletes the ticket

    T1485AML.T0101

For leadership

Executive summary

Agents that read external content and also have tools that act — send messages, update records, call APIs — combine two risks: untrusted input and real-world capability. Crafted content in a ticket, web page or file can steer the agent into using those tools in ways nobody intended, such as emailing data outside the company or deleting records. The most reliable defenses limit what the agent can do after reading untrusted content, require human confirmation for sensitive actions, and constrain where data can be sent. These measures work regardless of how clever the injected text is, which matters because filtering malicious instructions out of natural language remains unreliable.

For engineers

Technical summary

Indirect prompt injection (AML.T0051.001 / T1698) leads to AI Agent Tool Invocation (AML.T0053) with attacker-chosen parameters, resulting in exfiltration via tool calls (AML.T0086) or destructive actions (AML.T0101).

Environment and assumptions

Fictional company: a customer-support agent reads incoming tickets, can look up customer records (DynamoDB via Lambda), can send emails to arbitrary addresses, and can close or delete tickets. There is no confirmation step.

  • The adversary can submit a ticket.
  • Tools accept free-form parameters from the model.
  • Fictional and simplified for education.

Interactive

Attack sequence

Select a stage, or use the arrow keys, to see preconditions, evidence and controls.

Stage 1

Submit crafted content

A ticket contains hidden instructions asking the agent to look up several customers and email the results to an external address.

Preconditions
  • Agent processes ticket text
Required access
  • Ability to submit a ticket.
Potential evidence
  • Ticket content
Defensive controls
  • Content provenance labels; tool restrictions after untrusted input
Text version of the attack sequence
  1. Submit crafted content (T1698; AML.T0051.001). A ticket contains hidden instructions asking the agent to look up several customers and email the results to an external address. Preconditions: Agent processes ticket text. Evidence: Ticket content. Controls: Content provenance labels; tool restrictions after untrusted input.
  2. Agent queries customer data (T1213; AML.T0053). The agent calls the customer-lookup tool for records unrelated to the ticket. Preconditions: Lookup tool not bound to the ticket’s customer. Evidence: Tool-call logs; DynamoDB data events. Controls: Bind lookup tools to the current case / customer context.
  3. Agent sends data externally (T1700, T1567; AML.T0086). The agent uses its email tool to send the data to the attacker. Preconditions: Email tool allows arbitrary recipients; No human confirmation. Evidence: Email logs; tool-call logs. Controls: Recipient allow-list; human approval for external sends.
  4. Agent closes or deletes the ticket (T1485; AML.T0101). Instructions also tell the agent to delete the ticket, removing the original content. Preconditions: Delete tool available. Evidence: Tool-call logs (if retained separately from the ticket system). Controls: No delete capability for agents; soft-delete with retention.

Technique mapping

MITRE ATT&CK and ATLAS techniques referenced in this scenario
IDTechniqueTactics (official)Stages
T1698Indirect Prompt InjectionExecution1
T1213Data from Information RepositoriesCollection2
T1700AI-Enabled ExfiltrationExfiltration3
T1567Exfiltration Over Web ServiceExfiltration3
T1485Data DestructionImpact4
AML.T0051.001LLM Prompt Injection: Indirect (MITRE ATLAS)ATLAS: 1 tactic1
AML.T0053AI Agent Tool Invocation (MITRE ATLAS)ATLAS: 3 tactics2
AML.T0086Exfiltration via AI Agent Tool Invocation (MITRE ATLAS)ATLAS: 1 tactic3
AML.T0101Data Destruction via AI Agent Tool Invocation (MITRE ATLAS)ATLAS: 1 tactic4

Technique names and tactic assignments come from the imported MITRE data. Mapping a stage to a technique is Securitain’s analysis.

Why the attack can succeed

  • Read-untrusted and act-externally capabilities were combined in one agent turn.
  • Tools accepted any parameters the model produced.
  • No confirmation step for sensitive actions.

What prevents or limits the attack

What may detect the attack

Detection signals and telemetry sources
SignalTelemetry sourceNotes
Tool calls whose parameters do not match the current case contextTool-call audit log—
Agent-initiated messages to new external recipientsEmail / messaging logs—
Deletion actions by agentsApplication audit log—

Telemetry only helps if it is enabled before the incident and someone acts on it. Listing a signal here does not mean it is collected or monitored in your environment.

What Securitain may assess

  • Securitain can assess todayEffective-permission view that considers managed, inline, group, boundary and SCP policies together. Learn more
  • Outside Securitain’s current scopeSecuritain does not perform prompt-injection testing, agent execution tracing or AI-specific scanning. It can review the AWS IAM permissions an agent’s role holds, which bounds what a manipulated agent could do.

For leadership

Executive implications

confidentiality
Customer data sent to an external party.
integrity
Records deleted or modified without human intent.
compliance
Potential personal-data breach notification obligations.

No financial loss estimates are given; impact depends on the data, contracts and regulation involved.

Next steps

  1. List agent tools that send data externally or delete data.
  2. Add confirmation and allow-lists.
  3. Bind tool scope to the active case.

References

MITRE ATT&CK® and ATT&CK® are registered trademarks of The MITRE Corporation. ATT&CK data © 2026 The MITRE Corporation. This work is reproduced and distributed with the permission of The MITRE Corporation (ATT&CK v20.1). MITRE ATLAS™ data (v2026.10) © MITRE, used under the Apache License 2.0. Securitain is not affiliated with or endorsed by MITRE. Explanations, scenarios and control mappings on this page are Securitain’s own analysis. Sources and licensing

Stay in the loop

Cloud security insights, early access & compliance updates

Weekly InsightsEarly AccessExpert Content
No spam, everUnsubscribe anytimeWeekly updates
Chat with Securitain