The membrane
AI you can trust with your brand, because every decision it makes is on the record.
AI now writes code, installs software and talks to other systems on your behalf. The membrane is how we keep that safe: nothing important changes without a person's say-so, every automated decision leaves a receipt, and we publish our own scorecard, including what we got wrong.
- The promise
- Who we protect
- How we secure
- Data quality
- Give vs take
- How it works
- Protection, layer by layer
- Our scorecard
- Compare vendors
- AI cost and value
- Works with your tools
- Accountability
Three promises you can check
A person approves what matters
Releases, deletions, publishing and anything that can't be undone wait for a named person's go-ahead. Thousands of routine decisions run automatically; the irreversible ones never do.
Every decision leaves a receipt
Each automated decision is written to a ledger: what was decided, why, and by what. When a regulator, an auditor or your board asks "how do you know," the answer is a record, not a reassurance.
We publish our own scorecard
Our numbers are measured, dated and sourced, and when one turns out wrong we correct it in public with the old number printed beside the new one. Nothing here claims 100%.
Who we protect
Teams without a security army
Small security teams, managed service providers and growing companies who need enterprise-grade protection without enterprise headcount or price.
Companies putting AI to work
If AI assistants write your code or connect to your systems, we check what they install and what they connect to before it happens, not after.
Anyone who has to answer for it
Leaders who will be asked "how do you know we're safe?" get a record they can show, not a vendor's promise.
How we secure
Before it gets in
Software packages, AI tools and the instructions AI follows are checked before they're allowed in. Known-bad is refused with the reason; anything uncertain goes to a person.
At the front door
Our threat feed blocks known attacker infrastructure at the edge of the network, for us first and for every customer who uses it.
Watching ourselves
Dozens of automated checks re-measure our own systems every day and flag anything that drifted, because the failures that hurt most are the ones nobody is looking for.
How we make sure the data is good
Every record is verified
Nothing counts as stored until it's read back. Every indicator carries a confidence score and its source, so your team decides how strict to be.
Guardrails against our own mistakes
Addresses that can never be an attacker, like your own router or public DNS, are filtered out automatically, so our feed can't block your business by accident.
Corrected in public
When we're wrong, we say so on our updates page, old number beside the new. We credit every researcher whose work we use and never present someone else's findings as ours.
We give far more than we take
This is the point. Most security vendors take your data to build their product. We built ours to need almost none of yours.
What we take
An email address and organization name to issue your key. When you check a package, its name, ecosystem and version. Usage counts and a hashed network address, so we can keep the free tier free of abuse.
What we never need
Your code, your prompts, your customers' data, or your logs. Our checks work on public facts about software and infrastructure, not on anything private of yours.
What we give
Verified threat intelligence, the checks that run before anything gets in, open-source integrations, and our methods and corrections in public. Live numbers below.
How it works: a loop that checks itself
We see threats, verify them, ship protection, watch it work, and learn from what we got wrong. Each turn is measured. Click a stage for its live number.
Protection, layer by layer: from the developer's laptop to your network
Think of the business as a living thing with layers: where code is written, where it's built, where it runs, where it meets the internet, and where it protects you. Every plate is a real safeguard that exists in our code. The red ring is AI, held in check at every layer. Drag to turn it; click any plate.
The six checkpoints, live this week: what crossed each one and what happened to it.
Our scorecard: how fast and how safely we ship
DORA, Google's DevOps Research and Assessment program, is the industry's standard yardstick for software delivery. Its yearly survey sorts teams into elite, high, medium and low performers using four measures. In the 2024 report about one team in five reached elite. Here are ours, measured from every release we made in the last week, failures included.
| Measure | Ours | Elite benchmark | Where we rank |
|---|---|---|---|
| Deployment frequencyHow often we ship to production | 9.4 a day | On demand (several a day) | Elite |
| Lead time for changesFrom finished code to running in production | 5 minutes (median) | Less than a day | Elite |
| Change failure rateChanges that hurt users | 1.9% (1 of 54) | About 5% or less | Elite |
| Time to restoreWhen a change does fail, how fast users are whole again | Under 1 minute | Less than an hour | Elite |
What this means: we ship small changes many times a day, each one checked automatically after release, and the one change last week that would have affected users was rolled back by the system in under a minute. Speed and safety aren't a trade-off here; the checks are what make the speed safe. For full transparency: 5 of 54 release attempts (9.3%) failed their checks before reaching anyone and were stopped there.
Benchmark: DORA 2024 Accelerate State of DevOps performance clusters, as reported. Uptime from our availability monitors isn't quotable yet; they began recording on October 15, 2026.
Questions to ask any AI security vendor
We'd rather give you the questions than a chart about competitors we can't verify. Ask every vendor these, including us.
| Question | Our answer |
|---|---|
| Can I see the price before I talk to sales?A 2026 pricing-transparency study found no dedicated AI-security startup publishing prices on its own site. | Yes. Free, $24, $99 and $995 a month, on our pricing page. |
| What of my data do you need?Tools that sit inline inspect your prompts, responses or traffic to do their job. | Package names and versions you choose to check, and an email for your key. Never your code, prompts or customers' data. |
| Can I start without a contract? | Yes. A free key in about 30 seconds, no card. |
| How much access do you need before I see any value?Tools that inspect traffic must sit in its path; tools that scan your environment need read access to it. Ask what you grant on day one. | None beyond an email address. Every further layer is optional and asks only for the one permission it uses, like adding a blocklist to your firewall. |
| Is what you ask for proportional to what I get? | Each permission maps to one protection you can name, says what it touches and how to undo it. Nothing broad "just in case." See Getting started. |
| Do you publish your delivery scorecard? | Yes, above, measured against the DORA benchmark, failures included. |
| Where do you publish your mistakes? | On our updates page, with the old number beside the new one. |
| Can my team read the code that touches our systems? | Yes. Our integrations are open source and public. |
| Who approves the irreversible steps? | A named person, every time, with a record of the decision. |
Use both. Prompt firewalls guard the conversation between people and models at runtime. The membrane guards what gets into your systems in the first place: the packages, AI tools and instructions your developers and their AI rely on, and the attacker infrastructure at your edge. Different doors, same house.
Pricing study: AI Security Pricing Transparency Benchmark (2026).
AI: what it costs and what it's worth
Every AI call our platform makes is counted and priced. Spending on AI is only responsible next to the value it delivers, so we show both.
Works with the tools your team already uses
Security platforms, firewalls, developer tools and approval workflows. Open source and free to inspect. Click one.
Accountability you can show your board
Consent before anything changes
The instructions our own AI follows can only change with two keys: the AI's written statement of what changed and why, and a person's private approval that the AI can't see or fake.
Records, not reassurances
Releases, approvals and safety decisions are kept in permanent records, failures included. A failure rate means nothing unless you count every attempt.
People where it counts
Routine decisions are automated so people can focus on the ones that matter. Anything irreversible waits for a human, and a message that merely says "go" never counts as approval.
Measured from live sources on 2026-10-15. 95% confidence cap: nothing here claims 100%. No SOC 2 report exists and nothing here claims one. What changed · API docs · dugganusa.com