Back to Product
H7 Threat Model — v1.1

What H7 protects against
— and what it does not

H7 operates at the kernel syscall layer. It is designed to detect behavioral anomalies in autonomous agents that are invisible to signature-based EDR, prompt-injection filters, and network monitors. This document defines the threat surface H7 covers and its explicit boundaries.

Attack vectors H7 detects

Primary attack vectors detected at the kernel behavioral layer.

Living-off-the-Land (LOTL)

Critical

Attacker uses only legitimate OS syscalls — already whitelisted by EDR — to pivot laterally. Invisible to signature-based detection and prompt-injection filters.

H7 →

H7 detects LOTL via behavioral sequence analysis, not individual syscall inspection. The structural fingerprint of legitimate-but-unexpected call sequences is flagged within seconds.

Runtime Structural Drift

High

An agent's runtime behavior deviates from its declared specification after deployment — whether from prompt injection, model update, or environment compromise.

H7 →

H7 continuously compares live syscall sequences against the agent's baseline. Any structural deviation — including subtle ones — raises a scored drift event before damage occurs.

Supply-Chain Agent Injection

High

A third-party agent embedded in a CI/CD pipeline or orchestration layer is compromised upstream, prior to deployment. The agent itself passes static analysis.

H7 →

H7 establishes a behavioral baseline at first execution. Compromised supply-chain agents will exhibit behavioral patterns inconsistent with the baseline — caught on first run post-injection.

Prompt Injection

Medium

Malicious instructions embedded in data processed by an LLM-based agent alter its behavior at inference time — bypassing intent-level guardrails.

H7 →

H7 does not operate at the inference layer. It operates at the kernel layer. If prompt injection causes an agent to execute unexpected syscalls, H7 detects the structural consequence regardless of the injection vector.

.cal attestation chain of trust

From agent startup to offline-verifiable forensic certificate.

01

Agent starts

H7 eBPF probe attaches to the process at execve. No code changes to the agent required.

02

Baseline window

H7 records the agent's expected syscall-sequence model during a configurable baseline window (default: 30s).

03

Continuous monitoring

Rolling behavioral comparison against baseline. Drift score computed per batch of syscall events.

04

Detection threshold crossed

When the behavioral divergence measure crosses the detection threshold, H7 emits a signed .cal certificate and raises an alert. Containment action is operator-initiated.

05

.cal certificate emitted

Ed25519-signed attestation bundle generated: agent identity, timestamps, full trace, divergence measure reading, and alert action.

06

Offline verification

Any party with the published Ed25519 public key can verify the .cal certificate — no network, no CA, no SaaS.

Full threat matrix, attack-tree diagrams, and verification test vectors — available in the h7-demo-kit repository and upon request.

What H7 does not cover

Inference-time content safety

H7 does not inspect LLM outputs, prompt content, or semantic intent. It is not a content moderation layer.

Network-layer threats

H7 monitors outbound connection destinations and egress call rates from AI-runtime namespaces (ADR-019). It does not inspect payload content, TLS handshakes, or DNS queries.

Windows environments

H7 relies on Linux eBPF. Windows Subsystem for Linux (WSL) has partial support — see the demo kit for compatibility notes.

Hardware-level attacks

H7 does not detect firmware implants, Spectre/Meltdown-class CPU attacks, or physical hardware compromise.

Test the threat model in your environment

The demo kit ships with replay scripts for each attack vector listed above. Run H7 locally and validate detection before committing to a production deployment.

Clone Demo Kit ↗Book a Pilot (DORA / AI Act)