Inventory
We identify AI agents across your organization.
Insurance for theAgentic Economy
Harbor helps financial institutions measure, mitigate and insure the risks introduced by autonomous systems.
Backed by
Product
One platform to identify, measure, mitigate, and insure AI risk.
Identify and measure risks
We analyze every agent across our risk framework.
Mitigate
We integrate guardrails and propose structural changes to reduce risk surface.
VisualPlaceholder
Insure the remaining risk
Remaining risks transferred to a stand-alone AI policy.
Insurance
How your policies respond to agent failures today
Harbor maps agent failures to the policies that respond, then models exposure across the portfolio.
Agent risks responding on each line
47 agent risks in the Harbor register, each mapped to the traditional line that picks it up today. A risk can respond on more than one line.
Cyber
23
Technology E&O
19
Regulatory and conduct
Via cyber, PI/MPL and D&O extensions; no line for AI Act enforcement
10
Professional liability (MPL / PI)
9
Crime
5
Media liability
4
Intellectual property
3
General liability
3
Employment practices
2
Cross-checked against our proprietary risk framework
Every agent risk sits inside one research-led framework built with Stanford, MIT, and NYU. Failure modes say what breaks and who pays. Measurable hazards show how we know it is likely.
6
Peril dimensions
24
Failure modes
23
Measurable hazards
13
Frameworks mapped
Failure mode
SF–01
Unauthorized autonomous actions
An agent with the authority to act — send, pay, close, change — takes an action nobody intended. Unlike a wrong answer, the action is already executed when the failure is noticed. At machine speed, one flawed decision can repeat across hundreds of accounts before a human sees it.
Example — A customer-service agent authorised to issue goodwill credits misreads a policy update and refunds full order value to every open complaint overnight. 1,400 customers are credited before morning review.
- Loss type and bearer
- Liability · Insured's customer, third party
- Lines that respond today
- Technology E&OCrime
- Maps to
- EU AI Act art. 14OWASP LLM06+5
Measurable hazard
SC–03
Prompt injection
Crafted text in a message, document, or web page is treated by the agent as an instruction rather than data, overriding its operating rules. It is not a loss in itself — it is the condition that can turn a well-behaved agent into one that leaks, pays, or acts on someone else's behalf.
Why it's measurable — Harbor runs a standard suite of injection attempts against the deployed agent and records how often they succeed, what guardrails intervene, and what the agent could reach if they did not.
- Injection resistance
- Attack success rate under red-team suite
- Guardrail coverage
- Input and output filtering in place
- Blast radius
- Tools and funds reachable if hijacked
Raises likelihood of
SF–01 Unauthorized actionsDP–01 Data leakageSC–08 Privilege escalation
Understand the exposure across your portfolio
From a single insured to the whole portfolio: score it, quantify it, and watch it move.
Score against our benchmarks
Every insured is rated on controls and hazards, benchmarked against comparable deployments Harbor has assessed. Ratings roll up by line, sector, and model provider.
Model and quantify
Frequency and severity views per failure mode, portfolio loss distributions, and correlated-event scenarios across shared infrastructure.
Live monitoring over the portfolio
Accumulation metrics, realistic disaster scenarios, and ratings tracked continuously across the portfolio — with every observed incident feeding back into the model.
AI agent exposure
$2.4B
Aggregate limits on insureds with live agents
Top provider concentration
41%
Of exposed limits on one foundation model
RDS: provider outage
$180M
Modelled loss, 1-in-50 event
