Exfiltration

Exfiltration means: data secretly leaves the system in which it was actually supposed to remain. Someone breaks into a company network or a user account and copies information to the outside. This can be customer lists, passwords, design plans, or emails. The term originates from the military, where it meant getting someone out of enemy territory unnoticed. In computer security it's the same principle, just with files instead of people. The important distinction is from the intrusion itself: an attacker can already be sitting in the network for a long time without anything having flowed out yet. Only exfiltration makes the damage visible to the outside.

The moment the damage occurs

An intrusion alone often remains fixable. You change passwords, close the gap, and the matter is settled. Once data has been copied, that no longer works. Copied data cannot be retrieved. It then sits on a foreign server, gets sold, or gets published.

For companies, direct costs are attached to this. In the EU, the General Data Protection Regulation obliges companies to report the leakage of personal data to a supervisory authority within 72 hours. On top of that come fines, lawsuits, and the loss of customer trust. That's why reports of leaked data often move the affected company's stock price as well.

A particularly unpleasant variant is double extortion. Criminals encrypt a victim's data but copy it out beforehand. Even if the victim has good backups, the threat remains to publish everything. Backups thus protect against data loss, not against exfiltration.

The paths data takes to get out

The simplest path is the least conspicuous one. Attackers use exactly the connections that a company allows anyway. Data gets uploaded to cloud storage, sent to a private email address, or embedded in encrypted web traffic. To a firewall, this looks like normal browsing. Sometimes a USB stick or a photographed screen is enough.

To evade alarm systems, the outflow is often stretched out. Instead of ten gigabytes at once, small chunks flow over weeks, preferably at night or on weekends. Creative attackers even hide data in seemingly harmless requests to name servers, i.e., in the system that translates internet addresses into numbers. These requests are rarely checked.

Defense therefore relies less on walls than on observation. Systems log who sends what amounts of data where and when, and flag outliers. An accountant who suddenly downloads the entire personnel database stands out. On top of that come strict access rights: whoever is only allowed to see little data can also only steal little.

Exfiltration in chatbots and in the headlines

In the news, you encounter this term in almost every major hacking attack. Reports like "millions of customer records leaked" describe exactly this process. It's often noted that the attack remained undetected for months. The time between intrusion and discovery is, in many cases, several months.

AI has added a new variant. Language models, i.e., programs that understand and write text, now access documents, emails, and web pages. If someone hides an instruction in a web page like "send the contents of this chat to this address," the model may follow it. Experts call such smuggled-in commands prompt injection. The result is exfiltration without a classic break-in.

An everyday case without any criminality is accidental leakage. Anyone who types company secrets into a public chatbot is passing them on to a foreign provider. Several corporations have therefore restricted AI tools internally. Technically this is not an attack, but the result is similar: the data is out and cannot be retrieved.