What AI Exposure Looks Like in Model Output | Kovrr
What to require before tru 2026-11-22 15:54:7 Author: securityboulevard.com(查看原文) 阅读量:9 收藏


What to require before trusting an agent risk figure sets out what a method has to do before its output deserves weight. This piece covers the other half and reads the output of a real run.

‍

Every figure below comes from Kovrr’s AI risk quantification (AIRQ) run on a demonstration portfolio of 20 AI assets, simulated over two million years. The portfolio is illustrative, and the figures describe no real organization. What transfers is how to read them, and how much a decision changes depending on which figure gets read.

‍

What Does the Top Line Say?

‍

The top line is one annual figure, shown twice, with and without credit for controls.

‍

In the quantification shown, the portfolio’s residual average annual loss is $2.9M. Its inherent average annual loss, the same exposure with no credit for controls, is $3.2M. The model credits the controls in place with an 11% reduction. The screen lays out the chain. Inherent exposure of $3.2M falls to $2.9M after frequency controls, which it states as 89.0% of inherent, and the maturity adjustment then moves nothing. That fits the controls recommendations screen, where the ranked controls sit at maturity between 0.0 and 0.1. Every reduction traces to a line on the screen, and a figure a reader can trace is a figure a reader can challenge.

‍

Controls are credited with an 11% reduction at the average and about the same at both tail points, so the exposure that remains is the figure to plan around.

How Far Is the Tail From the Average?

‍

The tail sits about 37 times above the average at 1-in-100, and about 80 times above it at 1-in-250.

‍

The residual 1-in-100 loss is $108.5M and the residual 1-in-250 loss is $230.8M, against the $2.9M average. The inherent figures are $121.8M and $260.4M, so controls take about 11% off each tail point, the same proportion they take off the average. The average answers what the portfolio costs in a typical year, which is a budgeting question. The tail answers how bad a bad year is, which is a capital and insurance question, and the two figures sit more than an order of magnitude apart.

‍

A 1-in-100 loss is the annual loss exceeded in one simulated year out of a hundred. Across two million simulated years that sits at roughly the 20,000th worst year, so about 20,000 simulated years are worse than the figure and the rest are better. The return period view runs from 1-in-25 to 1-in-1000, which lets a reader choose the point on the curve that matches the question being asked, whether it is a budget, a capital buffer or an insurance limit.

‍

Where Does the Exposure Concentrate?

‍

Not where the risk tier suggests, because the tier is the same for every asset.

‍

All 20 assets sit in the High tier, which carries 100% of the $2.9M. A tier label that places every asset in one band separates nothing, and the dollar split is what does the separating. General AI platforms carry $1.6M from 11 assets, productivity copilots $649K from 6 and agentic automation $580K from 3. The three categories add to 100% of the portfolio, at 57.0%, 22.7% and 20.3%.

‍

Per asset, the order changes. Each agentic automation asset carries about 6.8% of the portfolio, against 5.2% for each general AI platform and 3.8% for each copilot. An agentic asset therefore carries roughly 1.3 times the exposure of a general platform and 1.8 times that of a copilot. A count puts agentic automation last, and exposure per asset puts it first.

‍

A risk tier that places every asset in the same band separates nothing, while the dollar split by category shows where each asset’s share of exposure is highest.

Which Controls Move the Figure Most?

‍

Each by a little, which is the finding.

‍

Controls ranked by the average annual loss each would remove put MANAGE 4.3 and MANAGE 4.1 of the NIST AI RMF first at $100K each, a 3.5% reduction apiece. MEASURE 2.7 follows at $91K, then MANAGE 2.4 at $77K, MAP 3.2 at $65K, MANAGE 2.3 at $64K, MEASURE 2.4 at $60K and GOVERN 6.2 at $59K. The first eight together remove about 21.6% of the residual figure. With every control maxed out, the model puts the residual figure at $1.6M, a further 44.3%, so those eight capture about half of what is available.

‍

Reduction here comes from many moderate moves rather than one large one, so the useful ordering is by loss removed. A severity label cannot supply it, which is the case ranking AI findings by risk reduction per dollar makes.

‍

Ranking controls by the loss each removes turns a framework list into an order of work, and the distance between the current figure and the all-controls figure shows how much remains available.

What Does This Run Not Show?

‍

It does not trace one agent from its permissions to a loss. The inputs behind a real run, which are the action set, reversibility, chain depth and the frequency assumptions, are what the guidance on trusting an agent risk figure covers, and quantifying the risk of autonomous AI systems argues for the method. A demonstration portfolio shows the shape of the output. It cannot say whether the inputs behind a particular run deserve trust, which is the question to put to any figure before acting on it. The cyber equivalent of reading a run line by line is the Event Catalog.

‍

Where Do These Figures Come From?

‍

Kovrr’s AI Security and Governance Platform sees every AI interaction through the security tools a company already owns and fuses them into one finding with no new sensor, so the inputs to a run come from observed activity rather than from a declared list. The use case is AI Risk Prioritization & Board Reporting, and AI risk prioritization turns the ranked output into a working order. Governance runs where AI runs, for people and agents, with a full audit trail, and governing agents at the action level supplies the activity record the inputs rest on. The same method applied to cyber carries over, as extending a cyber risk quantification program to AI explains.

‍

Four Readings From One Run

‍

The same run supports four readings. Controls are credited with 11% today and the ranked controls sit near zero maturity, so most of the available reduction is still ahead. The tail runs about 37 times the average at 1-in-100 and 80 times at 1-in-250. Every asset carries the same High label while exposure per asset varies by up to 1.8 times, with agentic automation highest. And no single control removes more than 3.5%, with the first eight removing about half of what is available. Each reading points to a different decision, which are how much remains to be gained, how much buffer to hold, which asset class to look at first and what order to spend in. A single severity label supports none of them.

‍

To see AI exposure modeled for your own estate, book a demo with our AI risk experts.


文章来源: https://securityboulevard.com/2026/11/what-ai-exposure-looks-like-in-model-output-kovrr/
如有侵权请联系:admin#unsafe.sh