Exfiltration · ATT&CK v20
T1700 AI-Enabled Exfiltration
A Exfiltration technique. Getting the data out.
- T1700Technique ID
- ExfiltrationTactic
- 4Places taught
What the attacker does
The attacker uses prompt injection or other methods to make an enterprise AI agent send data out with the tools it was given: an email tool, a web request, a file upload. The agent's sanctioned access does the stealing.
How it shows up in the logs
- An AI agent process reading data outside its normal scope, then soon after calling a tool that can transfer data.
- A connection from the agent to a destination outside the approved AI and business endpoints.
- Agent logs showing a tool call the user did not request, right after the agent read external content.
Questions to ask in triage
- What data did the agent read, and where did it send it?
- Which content caused it?
- Which tools and permissions does the agent need: remove the rest.
Read T1700 on attack.mitre.org
Where you meet it in CyberSim
In the theory
In the cases
Part of the answer in 1 investigation, in these learning paths:
Which case is not said here: finding it is the case.
More Exfiltration techniques
- T1041 Exfiltration Over C2 Channel
- T1567.002 Exfiltration Over Web Service: Exfiltration to Cloud Storage
- T1567.004 Exfiltration Over Web Service: Exfiltration Over Webhook
Source: MITRE ATT&CK T1700, tactic TA0010.