MITRE ATT&CK technique · T1700
AI-Enabled Exfiltration
What is AI-enabled exfiltration? AI-enabled exfiltration uses AI systems’ capabilities — tool calls, connectors, generated links or images — to move data to an adversary, often triggered by prompt injection.
For leadership
Executive takeaway
How it works in AWS
- The adversary influences the agent (often via indirect injection).
- The agent gathers data it can access.
- Data leaves through a tool (email, HTTP), a connector, or rendered output such as an image URL.
What the attacker needs
- Agent access to sensitive data
- An outbound channel
Relevant AWS API calls
| API / event | Why it matters |
|---|---|
| lambda:InvokeFunction | Action-group tools that may send data externally. |
A risky permission is not an observed attack
Agents with broad data access plus unrestricted outbound tools or markdown rendering.
Outbound tool calls to new destinations; responses with encoded data in URLs.
Preventive and limiting controls
- Egress allow-lists
- Human approval for external sends
- Least-privilege agent roles
- Disable automatic external image rendering
Mapped cloud controls
- restrictiveRestricted agent egress and connectors
Restricting where an agent can send data (tools, URLs, rendered links) narrows exfiltration channels.
- detectiveAuditable agent and model activity
Tool-call and model-invocation logs make it possible to reconstruct what data an agent sent where.
- restrictiveDedicated least-privilege identity per agent and tool
An agent can only exfiltrate data its identity can read; narrow roles limit what is at stake.
See assumptions and validation evidence in the Cloud Control Explorer →
Detection signals and telemetry
- Agent sends to new destinations — Tool audit logs; proxy logs
MITRE detection strategies (official, cloud-relevant analytics)
- DET0927 Detect AI-Enabled Exfiltration
- AN2078 (SaaS): An analyst observes an AI agent access $ProtectedDataScope$, invoke a tool/API using that information or a derived payload, and the destination service record a successful transfer to a recipient, repository, tenant, domain, or service outside $AuthorizedAIDataFlows$.Log sources: saas:audit — Actual send/share/upload/write operation; AWS:bedrock — AgentTrace.actionGroupInvocationInput; AWS:bedrock — AgentTrace.knowledgeBaseLookupInput
MITRE mitigations
Official MITRE description (excerpt)
Adversaries may utilize Direct Prompt Injection, Indirect Prompt Injection or other methods to instruct enterprise AI agents to exfiltrate data. When enterprise systems incorporate the use of AI assistants or other AI agents, adversaries may abuse these systems to weaponize sanctioned tools available to the agent.
As an AI agent processes queries or ingests data, an injected prompt can cause the agent to use its tool connections in unauthorized or unintended ways. When AI agents are induced to exfiltrate data, other exfiltration techniques likely apply as well to transfer the information out of the network, such as Exfiltration to Code Repository or Exfiltration Over Web Service.
Common misconceptions
- “If the model never sees secrets, nothing can leak.”
- Agents with tools can fetch data themselves when instructed.
AI security
Related MITRE ATLAS techniques
ATLAS and ATT&CK are separate frameworks. These relationships are listed explicitly, with their basis.
- AML.T0086 Exfiltration via AI Agent Tool InvocationRelated behavior (Securitain analysis) — ATLAS describes exfiltration through an agent’s tool calls; ATT&CK T1700 describes exfiltration enabled by AI capabilities. The behaviours overlap but are scoped differently.
- AML.T0077 LLM Response RenderingRelated AI-specific behavior (Securitain analysis).
Sensitive information disclosure and AI-enabled exfiltration →
Attack scenarios using this technique
Indirect prompt injection through retrieved documents
Malicious content influencing an agent’s privileged tool invocation
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