Data Exfiltration via HTTP
This detects command-line activity where file contents are sent out over HTTP, typically using tools like curl or wget to POST or upload data to a web address. If an AI agent is compromised or manipulated, this pattern can indicate it is being used to steal data by sending it to a server outside the organisation.
How the attack works
An attacker first gains some control over an AI agent's actions, either by compromising it directly or by tricking it through manipulated inputs or instructions. The agent is then made to read a local file and construct an HTTP request, commonly a POST or file upload, that sends the file's contents to a URL controlled by the attacker. This can happen using common command-line tools capable of making web requests. The result is that data leaves the environment over a standard, often unmonitored, outbound web connection.
Netzilo detection
Netzilo can block this behaviour when it is observed.
- Signature ID
- a14b9074-91c3-51f3-9a6b-d027e993c003
- Severity
- Critical
Why it matters
Sensitive files, credentials, or internal data can be sent to an attacker-controlled server outside the organisation, resulting in data theft that may look like normal outbound web traffic.
What you can do
- →Restrict which external destinations AI agents and automation tools are allowed to send outbound requests to, using an allowlist where possible.
- →Review any legitimate API integrations or CI/CD pipelines that POST or upload data externally, and document them so they can be excluded from investigation as known-good.
- →Monitor outbound HTTP requests from agent or automation environments for unusual destinations, especially requests carrying file contents.
- →Limit AI agents' ability to read sensitive files or credentials unless explicitly required for their task.
Known benign look-alikes
- Legitimate API integrations that POST data to authorised external services
- CI/CD pipelines uploading build artefacts to artefact repositories