An IoT platform is an attractive target. It holds credentials for thousands of devices, it can send commands to equipment, and it often serves many customers at once. A network firewall in front of it stops some attacks, but many of the dangerous ones look like ordinary requests: a valid sign-in from an unusual place, an API key used in a new way, a user reading far more records than usual. Catching those needs a layer that understands the application itself.
This article explains how the AI firewall built into Hermetry works, and the design choices behind it.
What it watches
The firewall listens to the platform, not only to the network. It works with more than 1,000 kinds of security signals: sign-in attempts, permission checks, API key use, request patterns, changes to devices and rules, and more. More than 1,200 kinds of actions in the platform are audited, so the firewall can see who did what, to which tenant, and when.
Three layers of detection
Rules and detectors
Some attacks have clear signatures: too many failed sign-ins, requests that probe for known weaknesses, a sudden burst of exports. Rules and purpose-built detectors catch these quickly and predictably. They are easy to read and easy to test.
A behaviour model
Other attacks only show up as a change in behaviour. A user who normally reads a few dashboards suddenly lists every device in a tenant. An API key that has called one endpoint for months starts calling ten. A behaviour model learns what normal looks like and flags what departs from it.
An attack-class model
A second, independent model looks at suspicious activity and estimates what kind of attack it resembles. Knowing the class helps decide the right response: a credential-stuffing attempt calls for a different action than data scraping.
The two models are independent on purpose. Neither acts alone. A model's finding leads to action only after it has been validated, so a single model's mistake does not lock out a legitimate user.
Decoys catch what detectors miss
Some of the most reliable signals come from traps. The platform contains:
- decoy API keys that no legitimate client ever uses;
- canary records that no normal screen ever reads;
- trap routes that no part of the interface ever calls.
Touching any of them is a strong sign that someone is exploring the platform with stolen access or an automated tool. Decoys produce very few false alarms, because legitimate users have no reason to reach them.
Observe, suggest, act
Security teams need to trust a system before they let it act. The firewall therefore has three modes.
- Observe: it records what it would have done, and does nothing. Use this to learn how it behaves on your traffic.
- Suggest: it proposes actions for a person to approve.
- Act: it takes validated actions itself, such as slowing or blocking a source, and records them.
Never on its own on industrial devices
One rule does not change with the mode. The firewall never acts on industrial devices by itself. Cutting off a pump controller or a gateway because of a suspicious pattern could do more harm than the attack it is trying to stop. For industrial devices the firewall only suggests, and a person decides.
Every decision is explained
A block that nobody can explain erodes trust quickly. Every decision the firewall makes comes with an explanation: which signals it saw, which rule, detector or model raised them, how the finding was validated, and what action followed. That lets administrators review decisions, reverse mistakes and tune the system with confidence.
How it fits with the rest of the platform
The firewall is one layer among several. Tenant data is separated in the database with row-level security, so even a successful attack on one account cannot read another tenant's rows. Roles limit what each user may do. Device credentials are issued per device, with certificates where devices need them. The firewall watches across all of it.
What to ask of any platform
- Does it watch application behaviour, not only network traffic?
- Can you run it in observe mode first?
- Does it explain its decisions?
- Is there a firm rule against automatic action on industrial equipment?
- Are actions audited, and can you see the audit log?
Read more on our security page and the AI firewall page.