Find Your Camera Blind Spots Before the Real Night

Most camera blind spots stay hidden until an incident exposes them. Sightproof stress drills test cameras against rare events they have no footage for, so you can find those gaps earlier. This post explains how they work, and what changed in October 2026 so the results can be trusted.

The problem: you can't test for what you've never recorded

Most security teams judge a camera by the alerts it has raised. That tells you what the detector catches in the footage it has seen. It tells you nothing about the events your cameras have never captured, and the events that matter most are rare by definition: a fence climbed at night, a worker collapsed in an aisle, an exit blocked during a shift change.

You can't wait for these to happen and then check whether the system noticed. You also can't ask staff to stage them at 2 a.m. across every site. So the blind spots stay invisible until the night they cost you something.

A detector's behavior also depends on its configuration: the model, the thresholds, the camera angle and the lighting. A model that looks fine on a bench can behave differently on one specific gate camera. Whatever you test has to be the detector as that camera actually runs it.

How Sightproof stress drills work

Sightproof can now run stress drills that test cameras against rare events they have no footage for. A drill generates synthetic CCTV-style frames for rare security and workplace safety scenarios, such as a night fence climb or a worker down. Blocked exits are another scenario type. Sightproof is one engine with packs: the Security pack is available and the Workplace Safety pack is in early access.

The frames then run through the camera's detector exactly as configured. This is the important part. Whether that detector is zero-shot with OWL-ViT or a fine-tuned model of your own, the drill exercises the setup you actually deployed, not a generic stand-in.

Each drill produces a miss report showing which rare scenes that camera would not see.

Simulated means simulated

A drill is a test, and we treat it that way. Every drill frame is clearly marked as simulated. Drill frames never become alerts, search entries or training data. Your incident queue, your searchable evidence and your fine-tuning sets stay free of synthetic scenes.

Why a miss only counts when the object is really there

A drill scenario is counted as a miss only when a second vision check confirms the expected object was visible in the scene. Scenes where the object never appears are marked separately and excluded from the miss rate. The same rule appears in the product specification: a miss counts only when a second vision model confirms the object was in the scene.

Review what the camera actually saw

The report is only useful if you can check it. Drill frames captured during security drills are now stored in platform storage and viewable from the app. You can look at the frame, see what the camera was shown, and judge for yourself whether the miss is fair.

If a frame cannot be uploaded, that failure is recorded rather than hidden. A gap in the evidence shows up as a gap, not as a silent pass.

If a drill points toward retraining, Sightproof can fine-tune detection models on alerts your team has verified from your own cameras. Drill frames themselves are never used as training data.

Where drills fit with the rest of Sightproof

Drills address the question of what a detector would miss. The rest of the workflow handles what it did see. Sightproof raises severity-based alerts with an acknowledgement and escalation workflow, groups alerts into incidents and tracks resolution, and keeps audit logs for security events.

For the incidents your cameras do raise, Sightproof assigns each one to a responder, keeps an activity log with notes, and requires a resolution note to close. An incident co-pilot drafts a summary, timeline, likely cause, unknowns and proposed actions, with every line tied to a source. Nothing happens automatically: proposals run only after a person approves them.

There is also an optional second look on live alerts, where a vision model judges whether the box shows what the detector reported. It runs in shadow mode and is measured against your team's own verdicts before it can mute anything. An agreement card compares the AI's calls against human confirm and false-positive verdicts, and muting stays locked until enough alerts have been reviewed and rejection precision is high enough. The two features follow the same principle: a model's judgment should be measured and checked before anyone relies on it.

On privacy, Sightproof blurs faces in camera frames before they are stored. For teams with data-residency concerns, Sightproof can run inference on-premise so data stays on your network, or deploy models on AWS, Azure or GCP.

Know your blind spots before the real night

A quiet alert queue doesn't mean the cameras are seeing everything. It can also mean the detector isn't seeing the rare things at all. Drills give you a way to find out on your schedule rather than an intruder's.

If you manage cameras at a site where a missed event would be expensive, start with one camera that matters, such as a gate, a perimeter fence or an exit. Run a drill, review the frames, and see what the miss report says. Learn more about Sightproof and request a walkthrough of stress drills on your own cameras.