averithBook a briefing

Built for black hat.Working for you.

A custom model, fine-tuned on datasets no one else can get access to, attacks your systems the way real adversaries would. A second model fixes what it finds. Then the attacker checks the fix holds.

Continuous Adversarial Verification for banks, payments, crypto and insurers. 24/7, inside the scope you authorize.

AST-042Payout ownership bypassVerified closedRed attackers latch onto the cell and breach toward its core; green antibodies fire back and close the path. The prints register, and the mark closes, only when the fix holds. Illustrative, synthetic data.

Attackers won’t wait foryour next security test.

AI makes sophisticated attacks easier to repeat. Annual assurance cannot protect a business that changes every day.

Test. Report. Wait.

The report describes yesterday’s systems. New code, integrations and access arrive tomorrow.

Defend. Challenge. Keep learning.

Continuous code defense and adversarial verification, working against the same changing business.

Across one year, changes land every week. A periodic test checks once. Averith checks continuously.JANFEBMARAPRMAYJUNJULAUGSEPOCTNOVDECCHANGESANNUAL TESTSNAPSHOTAVERITH

In financial systems,a software flaw becomes a business event.

An access-control failure is not just a technical score. It can expose money, confidential information and the trust the business depends on.

Money moves.Payments and assets
Who can approve a payout, change a destination or cross an account boundary?
Information carries value.Identity and confidentiality
Who can reach customer records, financial documents or privileged business data?
Trust is operational.Continuity and accountability
What happens to service, regulatory obligations and customer confidence when a control fails?

Security teams need evidence of business impact, not another unprioritized list.

One model attacks.The other protects.

Two model roles run continuously. One finds the path in, like an attacker would. The other checks code and changing systems, and closes the path out. Then the attacker challenges every fix, independently.

Two strokes closed by a verification bar A construction study of the Averith A. The attack stroke and the defense stroke stand apart on one shared baseline and never touch. Only the hatched verification bar, seated between their inner edges at the angle of each stroke, closes them into the letter. Mode: geometric construction of the Averith mark (64-unit viewBox), scaled 6x: X = 280 + 6(x - 32), Y = 62 + 6(y - 8). Attack stroke M6 56 28 8h12L18 56Z; defense stroke M38 22 58 56H44L32 34Z; bar M24 40h12l6 10H19Z. Grid: the mark's own 2-unit module (12 canvas units); every stroke vertex lies on it, while the bar's ends do not: they are set by the strokes' inner edges. The real bar's slanted ends tuck 1.3 to 1.75 mark units under both strokes, so only its exposed face is drawn; its sides coincide with the inner edges, which the strokes own in the still. During motion the bar carries its own sides (data-part verification-sides) and hands them to the strokes at contact. Hatch H1 identifies the bar as the verifying element. Grid spans that lie under a final form edge sit in grid-under-attack, grid-under-defense and grid-under-verification: hidden in the still, where the form owns the edge, and faded out in motion as each part lands. Motion: attack slides 48 and defense 48 units along the baseline; the bar rises 36 units from resting on the baseline. Labels are names for an illustrative reading of the mark, not product claims. Attack Defense Verification
The mark, constructed: the A only closes when verification seats between the two strokes.
  1. Adversarial verification

    Attack

    • Investigate reachable weaknesses.
    • Chain them the way a real attacker would.
    • Establish the potential business impact.
  2. Continuous defense

    Defend

    • Check code and changing systems.
    • Prepare and action remediation.
    • Return each change for verification.
  3. Both, in register

    Verify

    • Challenge the fix independently.
    • Close it only when the path is gone.
    • Carry the lesson into the next investigation.

Small weaknesses, chained into money.

Averith chains business logic: how access rules, APIs and money movements combine. Several findings that each look minor can add up to access to privileged accounts or funds. That is the risk signature matching alone does not show you.

Three small findings, one route Three short route segments – a session that survives a role change, sequential payout identifiers, and a payout edit that checks the account but not its owner – each rise one step. Aligned end to end they form one continuous route from the entry to the payout, three steps up, which no segment reaches alone. Averith · business-logic chaining. Mode: conceptual diagram, illustrative only; no metrics or customers. Relationship: three short segments become one route only when their ends align, and the aligned route reaches a level none reaches alone. References: SYS-05 (single-line route with R-bends ending in a destination circle), EMP-03 (route measured against level guides; a missing link as structure), INT-06 (separate parts projecting into one whole by a shared rule). Construction: level pitch h = 64; every segment is the same step: run 40, R16 bend, riser 32, R16 bend, run 40 (width 112, rise h). Entry at level 0, payout at level 3, so the payout sits 3h above the entry and only the three aligned segments reach it. Joint ticks mark where adjacent ends must coincide; they are fixed targets, not part of any segment. Payout highlight marks the reached result. Animation (src/scripts/plots/chain.ts) translates data-part groups vertically from loose offsets into these positions; this file is the final still. 3h h Payout Session Identifier Owner check
Illustrative chain · synthetic example

Everything an attacker brings.Working inside your scope.

An assistant answers questions. Averith acts like an adversary: specialist agents probing an estate in parallel, chaining what they find, and handing every path to a defender that closes it.

Live simulation of a synthetic estate. Red agents attack and chain; green closes and verifies.

  • A custom model, fine-tuned on datasets no one else can get access to.

    Non-public and dark-web intelligence, historical hacks and exploit behavior, and our own research and outcomes. Shaped by how real systems fail, not by what is on the open internet.

  • Tooling that lets it act like an attacker, not just answer like one.

    Specialist agents investigate in parallel. A separately fine-tuned decision model reads the evidence for each vulnerability class and chooses the next move: test, chain, fix or verify.

  • Systems that keep it current, and keep it remembering.

    New exploits are matched against your deployed estate as they surface. With memory on, every investigation starts from what Averith already learned in your environment.

Not another scanner.Not a once-a-year test.

How Averith compares with how each of the usual options typically works. Individual products vary.

Averith compared with annual penetration tests, vulnerability scanners and general-purpose AI testing agents
DimensionAnnual pentestScannersGeneral-purpose AI agentsAverith
The brainA human team, for a fixed windowRules and known signaturesA general-purpose modelA custom model fine-tuned on datasets no one else can get access to
How oftenUsually once or twice a yearScheduled or continuous scansOn demand or on a scheduleAlways on, 24/7
What it findsWhat fits in the engagement windowMostly known vulnerability patternsExploitable bugs in what you point it atBusiness-logic chains into money and privileged access
Who fixes itYour team, from a reportYour team, from a ticket queueYour team, often with a suggested fixA defending model prepares and actions the fix, with your approval or within your permissions
Proof the fix holdsUsually the next testThe next scanA retest, often on requestThe attacking model challenges it, independently and continuously
What it keepsA reportScan historyVaries by toolCustomer-scoped memory of what worked, what was blocked and why

The exposure is financial.Not merely technical.

About $30 billion: aggregate potential financial exposure from vulnerabilities we found and reported, each weighed by what an attacker could reach.

Business-critical findings
100+
Technology environments tested
80+
Retained fixes verified closed
95%

Anonymized cases
Evidence under NDA

A live, unauthenticated path into business APIs on a national-scale estate. The same class of weakness at a top-5 bank, a top-10 exchange, a national payments rail and international processors.

PHP2Shell: identified and patchedin two minutes.

When PHP2Shell leaked, Averith identified and patched deployed environments within two minutes of the announcement. Intelligence, exposure and patch worked as one loop.

  1. Know the attack

    Track new exploits, attack patterns and non-public intelligence.

  2. Find the exposure

    Find which deployed environments the new issue reaches.

  3. Close the path

    Prepare and action the fix within agreed permissions.

  4. Prove it holds

    Challenge the fix, then keep what is learned.

See the risk. Follow the agents.Control the action.

  • Business impactSee the estimate and its basis, not just a severity score.
  • Agent visibilityFollow active agents, new specialists and their work.
  • Human controlInspect evidence before approving a change.

Human-in-the-loop by choice.Autonomous within agreed boundaries.

You choose where a person must approve, and where Averith can act without waiting for one.

Only what is requested and authorized An agent’s requested actions overlap the customer’s authorized scope, and only the shared region is allowed. A dashed approval step on the path into that region applies when supervised and drops away under pre-authorized autonomy; the scope boundary stays in place either way. Mode: conceptual diagram, after AGT-05 “Only the shared scope” (intersection logic only). Mapping: circle = the actions an agent requests; rounded rectangle = the customer’s authorized scope; hatched segment = their intersection, the only actions it may take. Requested circle centre (204, 220), R108. Scope 208 × 232 at (252, 104), r12. The scope’s left edge x=252 is the chord of the circle, y 123.253–316.747; it is drawn once, by the scope. Hatch H1, 12-unit edge pitch, clipped to the segment. Transfer arrow on y=220 ends on the scope edge. The dashed bar at x=208 is an approval step that exists in supervised mode and is absent under pre-authorized autonomy; the scope boundary is solid in both. Illustrative only: no metrics, customers, or product interface shown. Authorized scope Requested Approval When supervised Allowed
How each operating mode handles investigation, remediation, verification and accountability
StageHuman-supervisedAutonomous
InvestigationAgents investigate within the approved engagement.Agents investigate continuously within the approved engagement.
RemediationA person reviews the evidence and approves the proposed action.Agents action fixes within your pre-authorized permissions.
VerificationThe adversarial side checks whether the approved fix holds.The adversarial side checks the change and continues the loop.
AccountabilityHuman decisions remain part of the evidence trail.Autonomous actions remain part of the evidence trail.

Autonomy changes who approves an action. It does not remove the scope.

Security that remembers.Protection that gets smarter.

The advantage is current intelligence and accumulated understanding, not access to a generic model.

A lesson, carried forward A path found blocked at the end of one investigation is carried, unchanged in size and position, into the start of the next investigation, which begins from that lesson and continues. Mode: conceptual diagram. Adapted from AGT-03 (a context carried between agents without changing its form) and AGT-07 (one hatched cell kept as a checkpoint at the same allocation). Local unit u = 24 SVG units. Each investigation record is 10u x 64; the lesson cell is 4u (232 to 328) in both, so the two projections at x = 232 and x = 328 are exact. Investigation 2 is offset 6u right and 7u down so its first cell is the lesson cell of Investigation 1. Stop bar at the lesson cell centre (x = 280); dashed absent segment = the route beyond the block, named by the Blocked label. Hatch H1, 12-unit pitch, clipped analytically to the 8-unit top strip of each lesson cell, as in AGT-07. Motion: the lesson cell travels 7u (168) straight down along the projections. Investigation 2's outline is drawn as its start edge (owned by the lesson it starts from), its top and bottom runs, and its end edge; round caps reproduce the rectangle's round corners, so the record can extend from the lesson. While travelling, the lesson carries its own top, bottom and right sides (data-part lesson-2-sides) and hands them to the record when the record forms. Illustrative only: no product metric or behaviour beyond opt-in memory carrying a lesson into the next investigation. Investigation 1 Blocked Lesson Investigation 2 Blocked 4u
  1. A separate model decides what to do next.

    Finding a weakness is one task. Choosing the next action is another. A separately fine-tuned decision model focuses on specific vulnerability classes: what the evidence means, what matters next, and which test, fix or verification to run.

  2. Memory of working with you, not just the latest prompt.

    With memory enabled, Averith brings relevant lessons from earlier work into new investigations in your environment.

    What carries forward with customer memory
    AreaStarting coldWith Averith memory
    ApproachesRediscover setup and investigation patterns.Bring relevant operational lessons into the next investigation.
    ObstaclesRepeat approaches that were already blocked.Use prior constraints to inform the next attempt.
    GuidanceKeep working from an outdated lesson.Replace earlier guidance as new experience is recorded.

    Memory is opt-in and scoped to each customer. Prior lessons guide an investigation; they never replace fresh evidence or expand authorization.

Banks, payments, crypto and insurance.Africa and the Middle East first.

Built in South Africa, against some of the most demanding financial systems on the continent. If it works here, it travels.

  • BanksPayouts, account boundaries and privileged access.
  • Fintech and paymentsAPIs and money movement between partners.
  • Wealth platformsCustomer assets and confidential records.
  • CryptoCustody, withdrawals and exchange logic.
  • InsurersNow holding the identity, data and payments banks carry, as insurance and banking converge.

How an engagement starts and grows

  1. Authorized discovery

    Prove

    A scoped, authorized engagement produces concrete findings and evidence.

  2. Executive case

    Business case

    Evidence is translated into business impact for the executive buyer.

  3. Integration

    Deploy

    Forward-deployed engineers integrate Averith into your environment.

  4. Contract

    Always on

    An annual enterprise contract for a deployed 24/7 managed service.

  5. Growth

    Expand

    Add systems, business units and group companies as value is demonstrated.

Built by people who havedone this work.

Averith is built in South Africa by security and fintech practitioners who have spent their careers inside the systems they now protect: banks, payment platforms and exchanges.

We work with security leaders who need proof that matters to the business, and with partners who can help us bring it to more institutions.

For investors

Averith is raising to turn proven capability into enterprise contracts. Capital goes to two priorities:

  • 1Compute, model training, evaluation and infrastructure.
  • 2Forward-deployed engineers who integrate Averith inside clients and expand adoption.
Request the investor brief
One stroke attacks. One defends. The bar is the proof that holds them together.

See what an attacker can reach. Then watch it close.

A briefing walks your security and executive teams through how an authorized proof engagement works on your systems.