Data Loss Prevention & Egress

Data Loss Prevention is the control class with the widest gap between its promise and its delivery. The promise — stop sensitive data from leaving — is exactly what everyone wants. The delivery is frequently theatre: a tool generating alerts nobody actions, blocking legitimate work while missing real exfiltration, and providing the appearance of control that’s arguably worse than none, because it substitutes for the harder work that would actually help.

Why most DLP is theatre

The honest assessment, because pretending otherwise wastes real money:

• Detection is genuinely hard. Distinguishing sensitive-data exfiltration from legitimate use requires understanding content and context, and DLP tools are noisy — a false-positive rate that gets them tuned toward permissiveness until they catch little.
• Trivially evadable by a determined insider. Encrypt before sending, use a channel the DLP doesn’t watch, photograph the screen. DLP that a motivated person walks around is providing assurance it can’t back up.
• It blocks legitimate work, generating friction that builds pressure to weaken it, until it’s configured so permissively it catches nothing.
• It substitutes for architecture. Effort goes into watching data leave instead of not exposing it or minimizing what exists to leak — treating the symptom while the disease compounds.

This doesn’t make DLP worthless. It makes it a control with real limits that’s routinely oversold, and the precision reality of detection applies in full — a DLP tuned to catch everything catches nothing usable.

Where DLP genuinely helps

Used with realistic expectations, DLP has real value:

• Catching accidents. Its best case, and a common real loss: the employee who emails the wrong spreadsheet, uploads to a personal account by mistake, misconfigures a share. Most data loss is error, not malice, and DLP catches errors well.
• Enforcing clear policy on well-defined data — blocking credit card numbers to personal email is tractable precisely because the data type is unambiguous.
• Visibility into where data flows, which feeds classification and lineage even when enforcement is limited.
• A deterrent and a compliance requirement — sometimes the point is the documented control, and that’s a legitimate if modest goal.

The reframing: DLP is good at accidents and policy enforcement on clear-cut data, and poor at determined adversaries. Deploy it for the former, don’t rely on it for the latter, and don’t let it substitute for minimizing what can leak in the first place.

Egress paths that matter

DLP is one control on the broader egress problem — the same outbound-is-what-matters point from the network side, here for data:

• Endpoint — USB, personal cloud sync, screenshots, print. The device is where data meets the physical world and the hardest surface to control.
• SaaS — the sanctioned app with an unsanctioned export, or the shadow SaaS nobody approved.
• Cloud storage — the public bucket, the over-shared link, the cross-account copy.
• AI tooling — the fast-growing and least-controlled path. Employees paste sensitive data into external LLM tools, and it leaves the moment they hit enter — a genuinely new egress channel most DLP wasn’t built for, and governing it is an unsolved corner.

The unifying principle is the network one: egress control is the direction that catches exfiltration, and it needs default-deny thinking rather than default-allow-and-watch. Watching every exit is harder than closing the ones you don’t need.

Insider risk and the surveillance tradeoff

Insider risk is where data protection meets a genuine ethical and practical tension, and it’s worth being honest about rather than waving through. Detecting insider data theft requires monitoring insiders — watching what employees do with data — and that monitoring is itself a privacy and trust problem:

• More monitoring catches more, and erodes trust and privacy — of your own employees, who have their own rights and their own regulatory protections.
• Surveillance has real costs — culture, morale, legal exposure in jurisdictions that restrict employee monitoring, and the corrosive effect of a workplace that treats staff as suspects.
• The tradeoff has to be made deliberately, with proportionality, transparency where required, and genuine limits — not maximized by default because the tooling permits it.

This is the same detection-versus-cost balance as everywhere in security, with the sharpest ethical edge, because the subject is your own people. The mature position is proportionate monitoring focused on genuinely sensitive data and high-risk actions, with transparency and limits — not a surveillance apparatus justified by a threat that’s real but rarer than the monitoring it’s used to license. Identity-centric signals — a dormant account waking, access wildly outside a role — are higher-signal and less invasive than watching everyone equally.

Where this connects

DLP and egress are the data-side of the network egress problem, bounded by the same precision limits as detection. The AI-tooling path ties to data governance for AI, and insider monitoring is threat hunting with the surveillance tradeoff made explicit — best served by the identity signals that catch misuse without watching everyone.