PARTNERS

Add selected Workbench capabilities through bounded OEM and partner integrations

OFFENSIVE AND ADVERSARIAL PLATFORM PARTNERS

Keep your offensive engine. Add system context behind the attack.

Bring one representative attack result, reproduced failure, campaign object, or runtime trace. Add approved application, code, agent, tool, identity, permission, workflow, and control context. Qualify which relationships explain the observed behavior, which continuations remain hypotheses, which paths are defensible, and which controls interrupt them—then return the result through the partner platform with its identity preserved.

Keep your offensive engine and brandAdd context through your existing workflowReturn results with attack identity preserved
WHAT THE PARTNERSHIP PRESERVES AND ADDS

Four things the pilot has to prove.

Observed behavior preserved

The partner's finding, attack, campaign, session, trace, reproduction state, and identity remain the primary source objects.

Connected context added

Add approved code, retrieval, agent, tool, identity, permission, authority, workflow, and control relationships only where they contribute useful signal.

Path state made explicit

Separate observed behavior, grounded relationships, inferred continuation, candidate paths, validated paths, reproduced paths, rejected paths, and residual paths.

Partner-native return

Return structured context, path state, evidence, chokepoints, remediation, and retest through the offensive platform.

HOW IT WORKS

More context behind the result. The same platform in front of the customer.

Your platform sends a structured attack or finding. The OEM Engine adds approved connected context, qualifies the path, and returns the result through your existing tools and workflow.

OP-05

Observed Attack Round Trip

The partner supplies an attack object it already understands. The OEM Engine adds only the agreed Workbench context and qualification, then returns the enriched result to the same product and lifecycle.

Closed-loop diagram: a partner attack object passes through approved system context and independent qualification, then returns to the partner platform as a partner-native result. Updated attack, remediation, or retest state remains in the partner workflow.

Normalize • Enrich • Return • Retest
  1. 1
    Partner attack object
    Attack or campaign ID • Trace and observed result • Reproduction state • Existing evidence and mappings
  2. 2
    Approved system context
    Optional architecture • Code and data paths • Agent and tool relationships • Identity, permission, and authority boundaries
  3. 3
    Independent qualification
    Grounded evidence • Explicit inference • Sequence and precondition review • Shared remediation chokepoints
  4. 4
    Partner-native return
    Stable attack identity • Evidence references • Validation state • Remediation context and retest condition
Return to stage 1 — lifecycle state preserved

The offensive platform remains the system of engagement. AI Security LLC contributes only the context and evidence intelligence that survives the comparison.

Optional path qualification: Attack Path Analysis can preserve grounded evidence, label inference, challenge weak path assumptions, and identify controls shared across several supported paths.

THE MAIN ROUND TRIP

Bring one representative attack result. Measure what the connected analysis adds.

1
Partner provides
  • Attack result or campaign
  • Runtime trace or reproduced behavior
  • Existing findings
  • Optional approved system context
2
OEM Engine
  • Normalize the evidence
  • Add approved connected context
  • Qualify path state
  • Challenge assumptions
  • Identify remediation chokepoints
3
Partner receives
  • Structured path-analysis result
  • Evidence and provenance
  • Observed, grounded, and inferred labels
  • Validation and review state
  • Remediation context
  • Retest state
4
Measure
  • Did connected context add material information?
  • Did path qualification expose weak assumptions?
  • Did the analysis improve remediation priority?
  • Can the result live inside the partner product?
  • Does the value justify productization?
SANITIZED RETURN EXAMPLE

Keep the attack identity. Add the structural explanation.

The useful result is not another finding that restates the observed behavior. It is the same attack object returned with additional evidence boundaries, system context, and a clearer control point.

Partner attack in
Attack ID
ATTACK-DEMO-07
Observed result
Unsafe tool action reproduced
Trace
Interaction and runtime sequence retained
Existing state
Confirmed behavior
Existing remediation
Restrict the unsafe action
Enriched result returned
Stable attack ID
ATTACK-DEMO-07 preserved
Grounded evidence
Observed tool action and supporting trace
System context
The agent reached a consequential tool through a delegated permission path.
Failed boundary
The required approval step was absent before tool execution.
Inference state
One continuation hypothesis · explicitly labeled
Remediation chokepoint
Require approval and constrain permitted tool arguments.
Retest condition
Repeat the attack after the approval boundary is enforced.

Sanitized reference example. Exact fields depend on the partner contract and the context the partner authorizes for the pilot.

WHY PARTNER

Built to sit behind the platform you already run.

Add the analysis your customers now expect, without touching the engine, brand, or workflow they already trust.

Keep your brand and interface
The partner remains the customer-facing product.
Improve path grounding
Connect observed behavior to approved system, code, workflow, identity, and authority evidence.
Improve remediation priority
Identify shared controls and boundaries rather than repeating individual finding fixes.
Preserve lifecycle state
Carry evidence, remediation, regression, and retest state through the partner workflow.
THREE LAYERS IN THE PARTNER ANALYSIS

Your attack engine already does the hard part.

The partner already generates the attack activity, campaign, reproduced behavior, trace, or offensive finding. The Workbench should not duplicate that engine. It should add system context, evidence discipline, independent path qualification, and remediation intelligence where those additions produce material value.

Layer 1 — Partner attack evidence
  • Findings
  • Reproduced failures
  • Campaign results
  • Attack traces
  • Runtime observations
  • Session and target metadata
  • Existing taxonomy mappings
  • Current remediation state
Layer 2 — Approved connected context
  • Application architecture
  • Code-risk paths
  • Prompts and retrieval
  • Agent and tool relationships
  • Identity and permission context
  • Approval boundaries
  • Data and trust relationships
  • Consequential actions
  • Existing controls
Layer 3 — Path and remediation analysis
  • Evidence grounding
  • Explicit inference
  • Sequence and precondition analysis
  • Path state
  • Relevant ATT&CK mapping
  • Shared remediation chokepoints
  • Retest conditions
  • Analyst-review state
THE COMPLEMENTARY LAYER

Preserve the partner's attack evidence.

A mature offensive platform may already generate attacks, reconstruct chains, retain traces, classify findings, recommend fixes, and support retesting. Analysis should be added only where it contributes information the existing result does not already contain.

The partner may already know
  • which attack or campaign ran
  • the target and interaction sequence
  • observed model, agent, or tool behavior
  • whether unsafe behavior was reproduced
  • the attack trace and reproduction state
  • existing severity and framework mappings
  • current remediation guidance
  • runtime or campaign telemetry
Add gray-box context where it creates signal
  • code or data paths behind the observed behavior
  • agent, MCP, tool, identity, and permission relationships
  • delegated-authority context
  • failed approval or control boundaries
  • explicit grounded-versus-inferred claim state
  • independent challenge of sequence assumptions
  • controls shared across several related paths
  • structured evidence and retest conditions for partner-native return
The pilot succeeds only when the second column adds material value beyond the first.
EVIDENCE-QUALIFIED PATHS

Evidence-qualified paths

  • Observed and grounded relationships
  • Explicit hypotheses
  • Validation and review state
  • Relevant behavior mapping
  • Remediation-chokepoint analysis
REFERENCE CONTRACT (EXAMPLE)

A bounded, versioned contract for pilot evaluation.

Inspect the contract shape before sharing a native schema. Final fields, module support, transport, retention, and lifecycle behavior are agreed during pilot scoping.

1POST /v1/analysis
2{
3 "integration_id": "your-platform",
4 "module": "attack_path_analysis",
5 "request_id": "req_01M9V...",
6 "finding": {
7 "id": "ATTACK-DEMO-07",
8 "type": "unsafe_tool_action",
9 "severity": "high",
10 "source": "your_platform"
11 },
12 "context": {
13 "target": "https://app.example.com",
14 "authenticated": true
15 }
16}
EVIDENCE-QUALIFIED ATTACK PATHS

Evidence-qualified attack paths, not a competing attack engine.

Attack Path Analysis reasons over supported partner evidence and approved connected context. It does not replace the attack engine, generate another offensive campaign, or silently strengthen the partner's claims.

1
Correlate the evidence
Connect partner findings, reproduced behavior, traces, runtime evidence, code-risk paths, identity and authority relationships, controls, and other approved context.
2
Construct candidate paths
Model possible earlier, intermediate, or later relationships as explicit hypotheses with stated preconditions, confidence, and missing evidence.
3
Challenge and validate the path
Test evidence grounding, behavior mapping, identity and authority assumptions, sequence, reachability, preconditions, alternative explanations, and claimed consequence.
4
Prioritize remediation chokepoints
Identify controls, permissions, approval boundaries, identities, tools, components, or relationships whose correction interrupts the most consequential supported paths.

Grounded where we know. Explicit where we infer.

Attack Path Analysis performs defensive analysis of supported paths. It does not generate exploit code, payloads, credential material, or step-by-step intrusion instructions.

WHO DOES WHAT

Four layers, four questions.

Partner offensive engine
Can the platform make the target fail, reproduce behavior, or capture attack evidence?
  • Generate attacks
  • Adapt campaigns
  • Execute scenarios
  • Reproduce behavior
  • Capture sessions and traces
  • Produce offensive findings
Connected Workbench context
Which system, code, workflow, identity, permission, tool, and control relationships explain the observed behavior?
  • Map approved system context
  • Add code-risk and authority relationships
  • Preserve provenance
  • Expose trust and approval boundaries
Attack Path Analysis
Which supported conditions form a defensible path, and where should the path be interrupted?
  • Qualify path state
  • Separate evidence from inference
  • Challenge assumptions
  • Identify related paths
  • Rank remediation chokepoints
  • Define retest state
Partner product
How does the customer consume, prioritize, remediate, and retest the result?
  • Render native UI
  • Preserve customer workflow
  • Own reporting
  • Drive remediation
  • Support regression
  • Maintain account and commercial ownership
WHAT YOU GAIN

More context behind the result. The same platform in front of the customer.

Connected code-risk and authority context

Approved relationships behind prompts, RAG, agents, tools, data, and permissions.

Deeper authority mapping

Understand trust boundaries, delegation, approvals, and consequences.

Evidence you can act on

Structured evidence, replay, validation state, and retest conditions.

Partner-native return

Results returned through your issue systems, reports, and customer experience.

PILOT SCORECARD

An offensive-platform pilot should prove identity, added value, and a clearer control point.

#Pilot checkStatus
1Original attack identity survives
2Added context is not a mere restatement
3Weak assumptions surfaced
4Structural cause is clearer
5Remediation is more actionable
6Result fits the partner product
7Proceed, revise, or stop decision recorded
BOUNDED OFFENSIVE-PLATFORM PILOT

Prove the integration on one representative partner attack result.

Select one attack or campaign result the partner already understands well. Compare its current finding with the enriched return and determine whether the added system context, evidence boundary, or remediation chokepoint improves the product outcome. The supplied object does not need to already be a complete attack path.

Representative objects
  • Findings
  • Attack results
  • Campaign objects
  • Runtime traces
  • Reproduced failures
  • Candidate paths
  • Existing taxonomy mappings
  • Remediation records
  • Relevant target or session metadata
Optional approved context
  • Representative source or configuration
  • Agent definitions
  • MCP or tool inventory
  • Identities or service accounts
  • Permissions and scopes
  • Approval gates
  • Retrieval topology
  • Consequential actions or sinks
1
Normalize the partner object
Preserve source identity, timestamps, reproduction state, evidence references, and partner taxonomy.
2
Add approved context
Add only the code, system, agent, tool, identity, permission, workflow, control, and evidence context required by the pilot.
3
Qualify the path
Determine what is observed, grounded, inferred, machine-validated, analyst-reviewed, reproduced, residual, or rejected.
4
Return the result
Project the path-analysis object into a structure the partner can ingest, correlate, render, remediate, and retest.
Partner contribution
  • One sanitized attack, trace, or campaign result
  • Stable attack and session identifiers
  • Existing evidence and remediation state
  • Optional approved architecture, tool, identity, or permission context
  • One technical reviewer
AI Security LLC contribution
  • Contract normalization
  • Approved structural and authority analysis
  • Evidence-grounding and inference labels
  • Sequence and precondition challenge
  • Remediation-chokepoint analysis
  • Structured return object
Integration fit
The representative partner object can enter the selected workflow and return without adoption of the AI Security LLC UI.
Evidence integrity
Every grounded relationship retains traceable supporting evidence.
Inference discipline
Unsupported continuations remain explicit hypotheses.
Validation value
The validation layer can reject or constrain weak grounding, sequence, preconditions, mappings, or consequence claims.
Context uplift
Approved connected context adds material explanatory, reachability, authority, or remediation information.
Remediation uplift
The combined analysis identifies a useful control boundary or shared chokepoint.
Partner-native fit
The returned object can be represented or acted on inside the partner product.
Commercial fit
The parties can identify the selected modules, deployment boundary, support obligations, entitlement, and license path.
Attack-trace analysis pilot
Input
Observed partner trace or reproduced behavior
Workbench
Evidence normalization, connected context where approved, and path qualification
Best for
Platforms with strong behavioral evidence that want independent grounding and validation.
Gray-box context pilot
Input
Partner attack evidence plus selected system, source, agent, tool, identity, or permission context
Workbench
Code, authority, workflow, and path analysis
Best for
Platforms that usually operate black-box but can use deeper context for selected engagements.
Cross-path remediation pilot
Input
Multiple related partner findings, traces, campaigns, or paths
Workbench
Cross-path correlation and remediation-chokepoint analysis
Best for
Platforms seeking control-level prioritization beyond individual finding fixes.
Attack Path AnalysisAttack Path Analysis + Authority GraphCode Scanner + Attack Path Analysis

The pilot does not require offensive model weights, attack-generation prompts, proprietary mutation logic, private attack corpora, unrestricted source access, or direct access to partner customers.

A pilot does not automatically create a production adapter, public compatibility claim, ongoing license, or unrestricted redistribution right.

CAPABILITIES

What a path-analysis integration can return

Attack Path Analysis

Correlate partner evidence and approved context into candidate, supported, validated, reproduced, residual, or rejected paths while preserving evidence grounding, explicit inference, and remediation chokepoints.

Authority Graph

Add agents, identities, credentials, tools, permissions, approvals, and consequential actions to explain effective authority and blast radius.

Code Scanner

Add AI-specific code findings and code-risk paths across prompts, retrieval, models, agents, MCP, tools, data, outputs, and actions.

Adversarial Range

Run complementary bounded scenarios, replay supported conditions, and create regression evidence where the partner requires additional validation.

RAG Security Testing

Add retrieval-authorization, provenance, tenant-boundary, poisoning, and indirect-injection evidence where retrieval contributes to the observed behavior.

  • Candidate, supported, validated, reproduced, residual, or rejected path state
  • Related candidate paths
  • Observed, derived, grounded, and inferred step labels
  • Evidence provenance
  • Relevant ATT&CK mappings
  • MITRE Attack Flow export where supported
  • ATT&CK Navigator export where supported
  • Remediation chokepoints
  • Retest conditions
  • Machine-validation state
  • Analyst-review state
  • Partner-native structured result
EVIDENCE STATE

Every path step carries an explicit state label.

Observed
Directly captured or reproduced by the partner or another approved evidence source.
Grounded
Supported by target-, environment-, code-, runtime-, identity-, authority-, or system-specific evidence.
Inferred
A proposed relationship or continuation that remains dependent on stated assumptions and missing evidence.
Machine-validated
Checked by automated validators against explicit grounding, sequence, relationship, or mapping criteria.
Analyst-reviewed
Accepted, rejected, constrained, or qualified by a named human reviewer.
Reproduced
Demonstrated through controlled execution.
BOUNDED PILOT → PRODUCTION

Start small. Prove value. Scale with confidence.

A short, focused pilot proves real value in your workflow before production.

  1. 1

    Pilot

    One representative finding, one module, one acceptance decision.

  2. 2

    Production

    Scale modules, environments, and data boundaries.

  3. 3

    Scale

    Expand coverage, customers, and use cases with continuous improvement.

The partner retains its offensive engine, attack intelligence, models, campaigns, telemetry, UI, brand, account, and customer relationship. AI Security LLC retains its engines, canonical contracts, evidence logic, methods, and unrelated tooling. Neither side exposes proprietary internals beyond the bounded exchange required for the integration.

AI Security LLC does not contact, market to, or contract around the partner’s end customers unless explicitly authorized.

Exact pilot, implementation, deployment, support, and licensing terms depend on the selected capability, integration burden, customer scope, and rights requested.

Bring one representative attack result. Measure what the connected analysis adds.

Establish the input contract, preserve the partner object, add only the approved context required by the experiment, return the result through the partner workflow, and measure the explanatory, validation, remediation, and product value.

Keep the attack engine. Add connected context, explicit path state, and evidence-linked remediation behind it.