Skip to main content

SPINE · Economic Exposure

Evidence-Backed Economic Exposure. Better Business Decisions.

AI risk quantification you can take apart: SPINE Economic Exposure connects observed AI behavior and validated technical findings to the business processes they touch, and prices the loss as a range with its sources — assumptions, uncertainty, mitigation cost and remaining risk in view before anyone decides.Decision support only. External feeds are fixtures until keys and a licence review, with no calibration or back-testing yet. Every figure here is a synthetic example.

From evidence to consequence

A verified effect, followed to the business.

Exposure starts from something Mythos AI Security actually observed — never from a severity score with a dollar sign attached. Findings that enable the same loss cluster into one scenario, so nothing is counted twice. A scenario, not a forecast.
Synthetic scenario
  1. 01 · Verified effect

    An agent issued refunds above its approval limit.

    Source · Achilles effect ledger · Verified, reproduced twice

  2. 02 · Loss pathway

    Fraud loss through unauthorized refunds, plus operational recovery cost.

    Source · One of 15 loss families, with its loss components

  3. 03 · Exposure drivers

    How often it’s attempted, how many refunds pass, and their size.

    Source · Mythos-observed rate · customer-provided volume · expert estimate

  4. 04 · Possible response

    Bind approvals to amount and destination, then retest.

    Source · Remediation scenario · residual exposure modeled

Inputs, sources and validity

Every number shows its sources.

Thirty customer variables in nine domains each carry a source type, and every result a grade from A to D or Unknown — so a reader can see how much rests on customer data, observation or estimate.

Source types

  • Customer provided

    Volumes, values and process facts you supply.

  • Mythos AI Security observed

    Behavior and effects seen during testing.

  • Calculated

    Derived from other inputs, with the method recorded.

  • Official public

    Published public statistics.

  • Industry benchmark

    Sector reference figures, used with care.

  • Expert estimate

    Judgement, named and dated.

  • Unknown

    Missing — shown as missing, never filled in quietly.

Confidence grades

  1. AEvidence richMostly customer data and observation.
  2. BSupportedGood evidence with some estimation.
  3. CIndicativeLeans on benchmarks and estimates.
  4. DExploratoryMostly assumption; for discussion only.
  5. —UnknownCarries a banner and is excluded from totals.
Data dictionary
Valuation date
The date the result was true of.
Horizon
Annualized, with single-event loss kept separate.
Reporting currency
ISO 4217, with rate provenance; native amounts are never overwritten.
Model version
The model build that produced the result, so it can be reproduced.
Reviewer
A named non-author who reviewed the result before it issued.
Override
Applies only while in force; a sensitive one needs two other approvers.

Modeled exposure ranges

Ranges, never a single number.

100,000 seeded Monte Carlo draws give P10, P50, P90, the mean and a severe-but-plausible tail. Single-event and annual loss are kept apart; switch between them.

Refunds beyond approval · modeled loss range

Refunds beyond approval · modeled loss range, Single event: modeled ranges by band
BandModeled rangeWhat drives it
Low$0.1M – $0.4MRefunds caught within a day; small batch.
Base$0.3M – $1.2MTypical volume before detection.
High$0.9M – $3.1MSlow detection and large refunds.

Synthetic example · not a forecast

Remediation economics

What a fix could remove.

Current and residual expected annual exposure, the investment a fix needs, its payback and the tail it removes — decision support that never overrides a security veto.
Modeled loss range · one validated scenarioIllustrative
P10
$0.6M
P50
$2.1M
P90
$7.4M

Expected annual exposure

Today$3.2M
After the proposed fix$0.9M

A $0.4M fix that removes about $2.3M of modeled annual exposure — with the assumptions, sources and confidence shown beside every figure.

Two remediation options · synthetic example
MeasureTodayOption A · bind approvalsOption B · lower refund limit
Expected annual exposure$3.2M$0.9M$1.4M
Expected exposure reduction—$2.3M$1.8M
First-year investment—$0.4M$0.1M
Payback—≈2 months<1 month
Confidence gradeBBC

Option B costs less but refuses many legitimate refunds and rests on more estimation. Neither is a promised return; both are decision support.

Sensitivity and limits

See what the range rests on.

Move an assumption and watch the range respond. Sensitivity is shown apart from probability and from the confidence in the evidence — three different things.
Attempt frequency
Loss per event
Control in place

Synthetic example

Modeled annual exposure, central estimate

$3.2M

Range $0.6M – $7.4M

These are the scenario’s stated assumptions. Change one to see how much of the range rests on it.

A lookup into fixed synthetic results — not a model running in your browser.

Financial validity and receipt

A record a reviewer can take apart.

Each result leaves a signed Financial Exposure Receipt and a Validity Card: model version, sources, assumptions, unknowns, grade and valuation date. A non-author reviews it before it issues; an author never reviews their own scenario.

Financial Exposure Receipt · synthetic

Refunds beyond approval

Receipt
fer_2026_10_16_0412
Scenario
Refunds beyond approval · support agent
Result
Expected annual exposure $3.2M · range $0.6M–$7.4M (P10–P90)
Valuation date
16 Oct 2026
Confidence
B — Supported
Review
Non-author reviewer (role) · before issue
Show all 12 fields
Signature
Signed · verifies offline
Simulation
100,000 seeded Monte Carlo draws
Reporting currency
USD · native amounts retained
Model version
ee-model 0.3 (illustrative)
Sources
4 inputs · bound to the exact rows used
Scenario hash
sha256:9c41…e07a
Not a certification, a materiality finding or an actuarial estimate. The signature proves provenance, not that the figures are right.

Validity Card · synthetic

What this range rests on

Model version
ee-model 0.3 (illustrative)
System fingerprint
7f3a…c91e
Customer-specific share
55% of inputs
Benchmark-derived share
25% of inputs
Expert-estimated share
20% of inputs
Most sensitive variables
Attempt frequency; loss per event
Show all 11 fields
Unknown variables
Detection time after hours
Data freshness
Volumes as of September
Feeds
Fixtures until keys and a licence review
Calibration
Not calibrated or back-tested yet
Reproduction
seed 4417 · result hash recorded
A range is only as strong as its grade and the shares above it.

For the board and CFO

The questions a board will ask.

  • Does Mythos AI Security know what we would actually lose?

    No. Economic Exposure prices ranges under stated assumptions — P10, P50 and P90 from 100,000 seeded draws. It never predicts an exact loss, and it keeps single-event and annual loss apart.

  • Is this a guaranteed return on a fix?

    No. Remediation economics compare current and residual exposure against the investment a fix needs. They are decision support — not a promised return.

  • Does it decide whether something is material?

    No. It makes no legal-materiality determination and uses no unverified penalty figure. Materiality, disclosure and every business decision stay with you.

  • Where do the inputs come from?

    From your data, what Mythos AI Security observed, public sources, benchmarks and named expert estimates — each labelled and graded. Public feeds are fixtures until keys and a licence review; breach figures are priors, never multipliers.

  • Can we trust a range that rests on estimates?

    Only as far as its grade says. Each result shows how much is customer-specific, benchmark-derived or estimated, and which variables move it most. It is not yet calibrated or back-tested.

  • Is every figure signed and priced?

    Each scenario’s receipt is signed; the portfolio total is not itself signed yet. Harm to a person is never priced. Decision support only.

Mythos AI Security

Discuss exposure on one of your systems.

A scoped conversation about one AI system: what was observed, what it could mean financially, and what we would need to say it with confidence.