Harbor brings every AI system into one platform for continuous identification, measurement, and mitigation of AI risk.
Works with
Product Overview
Every session an agent runs, turned into dated evidence.
Harbor reconstructs each session from the agent’s own telemetry and replays it: against the rules you set, the limits you set, and the failure modes no rule names. Every detection is attributable to the session and the version that produced it.
Define what an agent is allowed to do, then enforce it at the moment of action. Harbor evaluates every tool call before execution, blocks anything outside policy, and records every decision as evidence.
Harbor maps the code paths, models, tools, and permissions used by an agent. Code analysis runs in isolated, ephemeral sandboxes, producing verifiable evidence while protecting customer code and data.
Harbor tests agents as a user or attacker would, probing for unsafe behavior, policy bypasses, and tool misuse. Tests and responses become evidence for assessing failures.
How it works
Input
Applicable risks
harbor risk registry
Harbor product
Agent version
cost if it happens
Artifact
Vulnerability reports
chance if it happens
Third-party
Agent Workflow
Harbor sandboxEnvironment
Harbor model gateway
Harbor product
Whitebox scan
prompts • tools • code
Skill
Adversarial probe
Blackbox • Attack library
Skill
Runtime evaluation
Trace • Imported findings
Skill
Triage and confirm
Reproduce • dedupe
Skill
Output
Confirmed findings
per agent version
Artifact
Severity
cost if it happens
Artifact
Severity
chance if it happens
Artifact
Deployment Methods
Fully Managed
Fastest to start, managed and upgraded for you. [time to first assessment] · shared tenancy, isolated data · [certification status]
Custom Control
Our control plane, your data plane. Traffic and traces stay inside your boundary.
Isolated & Controlled
Fully isolated cloud instances for regulated and air-gapped environments.
Harbor can be deployed to match the security and infrastructure requirements of the organization, from fully managed SaaS to hybrid cloud and isolated on-premises environments.
Product Metrics
Detection, attack success rate, latency, and throughput show not just whether the system works, but how reliably and efficiently it performs under production conditions.
41%
Detection
Classifier accuracy on a named, held-out eval set
50%
Attack Success Rate
Against what, using how many strategies
20ms
Latency
Against what, using how many strategies
20ms
Throughput
Against what, using how many strategies
Security
Independently audited security controls verified annually
Full compliance with EU data protection regulations
Integrate with your identity provider for seamless authentication
Fine-grained access control at the project and resource level
Full compliance with HIPAA requirements to secure PII
Deploy Harbor’s data plane on your own infrastructure
A risk framework and an insurance capacity, built for banks, insurers, and the teams that govern their models. Born out of research at Stanford. Built to sit on a bank’s balance sheet.
Pay small refunds out directly so customers are not waiting on agent