Back to Product
H7 Technical Architecture

What H7 observes,
and how it decides

H7 operates entirely at the Linux kernel layer via an eBPF probe. It does not read prompts, inspect model outputs, or intercept network payloads. It watches the structural behavior of agent processes — five detection channels, one cryptographic certificate.

The behavioral baseline

When H7 attaches to an agent process, it begins a 30-second baseline window. During this window, H7 records the agent's expected behavioral fingerprint across all five detection channels.

The baseline is nota list of allowed syscalls. It is a structural model of the agent's behavior — the distribution of scheduling events, the causal graph of process relationships, the set of network destinations it contacts, and the sequences of kernel operations it performs.

After the baseline window, H7 enters continuous monitoring mode. Every new batch of kernel events is compared against the baseline model. A deviation that crosses the detection threshold produces an alert and a signed .cal certificate — regardless of whether the individual syscalls are "allowed."

h7.sensor — baseline phase
[h7] probe attached → agent:vercel_agent (pid 8812)
[h7] baseline window: 30s
[h7] L1 collecting sched_switch distribution...
[h7] L2 recording process-execution graph...
[h7] L3 building causal relationship model...
[h7] L4 learning network destination set...
[h7] L5 monitoring ptrace surface...
[h7] BASELINE_ESTABLISHED ✓ (30.002s)
[h7] → continuous monitoring active

Five detection channels

Each channel operates independently. A single deviation on any channel produces an alert.

L1 · Behavioral entropy

Scheduling-entropy drift detection

H7 watches the kernel scheduler event stream for each monitored process namespace. It models the agent's expected scheduling behavior during a 30-second baseline window, then continuously compares the live event distribution against that baseline. A divergence measure is computed per batch; when it crosses the detection threshold, H7 emits an alert and a signed .cal certificate.

Observes

Scheduling event distribution across monitored agent namespaces.

Does not observe

Individual syscall content or argument values.

DRIFT_DETECTEDLIVING_OFF_THE_LAND
L2 · Shell spawn detection

Process-execution classification

When an AI agent spawns a shell interpreter (bash, sh, python, etc.) as a child process, H7 classifies the resulting process-execution event. Shell spawns are a canonical signal of prompt-injection and living-off-the-land attacks: the agent was instructed to run an OS command it was not designed to run.

Observes

Child process execve events from agent namespaces — specifically interpreter spawns.

Does not observe

The content of shell commands or arguments passed to the interpreter.

LLM_AGENT_HIJACK
L3 · Causal graph analysis

Structural process-causality analysis

H7 builds a causal graph of process relationships from the kernel event stream — which process opened which file, which process forked which child, which sequences of calls follow each other. Structural deviations from the baseline causal graph (unexpected directed cycles, unfamiliar call-graph patterns) are flagged independent of any single syscall.

Observes

Process-relationship graph derived from the kernel event stream.

Does not observe

File contents, network payloads, or memory contents of monitored processes.

STRUCTURAL_DRIFTCAUSAL_CYCLE_DETECTED
L4 · Outbound connection monitoring

Network egress detection

H7 hooks the kernel's outbound connect and sendmsg paths for AI-runtime namespaces. It monitors destination addresses against a configurable allowlist and detects egress call-rate bursts. An agent connecting to an unknown destination or sending at anomalous rates raises a network alert — without inspecting any packet content.

Observes

Destination address and egress call rate from monitored namespaces (ADR-019).

Does not observe

Payload content, TLS handshakes, DNS queries, or packet bodies.

UNKNOWN_DESTINATIONNET_EGRESS_BURST
L5 · Ptrace intercept

Process-attach detection

H7 monitors for ptrace ATTACH and SEIZE operations targeting agent processes. Process-attach is used by debuggers, memory scrapers, and lateral-movement tools to read or modify a running process's memory — a technique invisible to application-level security layers. Any unexpected ptrace operation on a monitored agent triggers an immediate alert.

Observes

ptrace ATTACH/SEIZE syscalls targeting monitored agent processes.

Does not observe

Legitimate debugger use outside monitored namespaces.

PTRACE_ATTACH_DETECTED

The .cal certificate

Every detection event produces one signed artifact. It is the only output H7 generates.

agent_idUnique identifier of the monitored process
ts_nsNanosecond-precision timestamp of the detection event
hostHostname of the machine running the H7 sensor
channelDetection channel (L1–L5) that triggered the alert
divergenceBehavioral divergence measure at time of alert
syscall_traceKernel event sequence preceding the detection
actionAlert action taken (ALERT_EMITTED; containment is operator-initiated)
signatureEd25519 signature over all fields — verifiable with the published public key
What it proves

That a specific agent process exhibited a specific behavioral sequence at a specific nanosecond-precision timestamp on a specific host — and that this record has not been modified since H7 signed it.

Who can verify it

Any party with the published Ed25519 public key — your auditor, your CISO, your legal team, or a regulatory body. No network access, no external CA, no SaaS dependency required.

Containment is operator-initiated

H7 detects, attests, and alerts. It does not automatically terminate or isolate agents. Containment requires an explicit operator action — preserving human control and audit trail continuity.

Architecture: zero external dependencies

All telemetry stays on the host. The Ed25519 public key is the only external dependency.

01

eBPF probe — on host

The H7 sensor runs as a Rust no_std eBPF program attached to the Linux kernel. Zero heap allocation. No kernel module. No root filesystem writes.

02

Behavioral analysis — on host

All detection computation happens locally. No telemetry is sent to a cloud service. No API call is required for detection to function.

03

.cal emission — on host

The signed certificate is written to the local filesystem. The Ed25519 signing key never leaves the host. Verification uses only the published public key.

04

Alert delivery — your infrastructure

H7 pushes alerts to a local sinkhole process. You integrate that with your SIEM, Slack, PagerDuty, or incident management system — H7 makes no external calls.

See all five channels detect in 10 minutes

The demo kit ships with attack scripts for L1, L2, L4, and L5. Clone it and run the detection cycle on your machine — no cloud account required.

Apply for a Pilot →Clone Demo Kit ↗