There's a gap.

You are the CISO.

Your stack: CrowdStrike on every endpoint. Zscaler proxying all outbound traffic. Proofpoint on email and DLP. A SIEM with 200+ custom detection rules your team spent 18 months tuning.

You are responsible for proprietary IP worth hundreds of millions in competitive advantage, customer PII across tens of thousands of accounts, infrastructure running across thousands of nodes, regulatory exposure under SEC, FINRA, and SOX, and a board that asks about security posture every quarter.

You have done everything right.

What follows is what your stack does not catch.

Situation 01: The trusted insider

A senior employee with access to everything decides to take it on the way out. Your security tools stopped every obvious move. He found one they could not see.

Senior quant researcher. Full legitimate access. Decides to take the codebase on the way out.

Attempt 1: USB drive

CrowdStrike caught it immediately. Falcon Device Control triggered within 800ms of insertion. The drive never mounted. Stack working exactly as designed.

Attempt 2: Email archive to personal Gmail

Proofpoint caught it before it left. DLP intercepted the outbound message. Attachment exceeded the 10MB threshold for external destinations. Recipient domain matched the personal webmail block policy. Stack working exactly as designed.

Attempt 3: Upload to Dropbox

Zscaler blocked before a single byte left. Cloud App Control categorized Dropbox as unauthorized personal cloud storage. Stack working exactly as designed.

Attempt 4: What your stack cannot see

He registers a private server under a generic hosting provider, names it like an internal dev tool, and writes a script that reads authorized files and sends them over HTTPS/443 to his server. Every file he touches, he is allowed to touch. The destination passes reputation checks. The traffic looks normal. There is nothing for your tools to flag.

• CrowdStrike: Authorized process. Authorized files. No malicious signature. No alert.
• Zscaler: HTTPS/443 to a domain with clean reputation. No policy violation. No block.
• Proofpoint: No email. No USB. No flagged destination. No alert.

Undetected. Every night for 6 weeks.

The result: six weeks. The firm's entire research codebase on a server he owns. He walks out. The algorithms walk with him.

Hilt would have stopped it. By day 6, Hilt had noticed this employee was reading far more files than he ever had, doing something with them at the same time every night, and sending an unusually large amount of data to a web address he had never visited. None of those things violated a policy. Together, they were completely outside his normal pattern. Hilt flagged it before week 2. All four attempts appear in a single timeline.

Situation 02: The contractor

An outside contractor brought in for a short project had limited access. Your tools caught his first two attempts. His third worked, because it used access he was supposed to have.

Third-party contractor. 3-week engagement. Read access to two database schemas.

Attempt 1: Personal cloud storage

Zscaler stopped it instantly. URL filtering blocked the request. Destination categorized as personal cloud storage, blocked firm-wide for contractor environments. Stack working exactly as designed.

Attempt 2: Reconnaissance tool

CrowdStrike caught it within seconds. Behavioral AI matched a known reconnaissance signature, killed it, quarantined the binary, escalated a high-severity alert. Stack working exactly as designed.

Attempt 3: What your stack cannot see

He stops breaking rules. Instead, he uses only the access he was given. He queries customer records he was technically allowed to reach, a little more each night, expanding quietly into data he had no business reason to access. He sends it out through a connection your team had specifically opened for him. No rules broken. No unauthorized tools. No blocked destinations.

• CrowdStrike: Normal database client. No signature. No alert.
• Zscaler: Internal traffic to provisioned systems. No block.
• Proofpoint: No email. No cloud upload. No alert.

Undetected. Three weeks. Full PII dataset gone through an authorized path.

Hilt would have stopped it. By day 5 it flagged three things that did not fit: he was reading customer records he had never accessed, his outbound data volume was much higher than usual, and the timing of his database activity matched his outbound transfers almost exactly, night after night. Not a policy violation. But not normal. Hilt flagged it.

Situation 03: The pipeline

This one had no attacker. A routine automated process had been given broad access during an emergency eight months earlier. Nobody cleaned it up. A configuration mistake meant it started putting sensitive customer data somewhere it was never supposed to go. Every night for six months with no alert, no flag, and no human involved.

A misconfigured ETL job. No human actor. No intent. Just permissions that were never cleaned up.

• CrowdStrike: No agent on the Lambda function. Nothing to instrument.
• Zscaler: Does not proxy internal AWS service calls. Out of scope by design.
• Proofpoint: No endpoint. No email. No user. No trigger.

180 days. No attacker. No alert. Found by an external auditor. By accident.

Hilt would have stopped it. On day 3, it noticed this job was writing to a location it had never used before, and that the data going there included customer records it had no reason to be handling. No human had to spot it. No audit. No alert rule had been written for this scenario. Hilt flagged it because it was new behavior for that process. Day 3. Not day 183.

The structural gap

Every one of these scenarios involves behavior that looks permitted at every individual step. The existing stack catches what it was built to catch: known signatures, policy violations, unauthorized destinations.

What it does not do is model what is normal for every process, identity, and data flow over time, or correlate the pattern across cloud, SaaS, endpoints, and AI agents simultaneously. That is the structural gap. Hilt closes it.

How Hilt works

Instead of starting from a list of known bad things, Hilt builds a model of what normal looks like for every person, every automated process, and every system in your environment, and watches for anything that deviates from that, even when it does not match any known threat.

It watches data movement at the kernel, where the move actually happens, so the responsible process does not have to be one Hilt was told about in advance. It runs entirely inside your own cloud, so none of your data goes anywhere, and it works on metadata by default, so it does not have to read your data. It gets more accurate over time: the longer it watches, the better it understands what normal actually looks like.

Hilt resolves each move to a real identity and the job behind it, where the source allows, and surfaces the dangerous pattern across moves. When a pattern is dangerous, it isolates the host at the network (quarantine) from the control plane. It stays off the path and never sits inline, so it never drops, filters, or alters your traffic.

The footprint is small and checkable: roughly 0.1% of one core and 4 to 8 MB of memory per host. The collector installs with a single command, no kernel module to compile and no reboot, reversible per node. It deploys entirely inside the customer's own cloud, and events never leave the account.