Add selected Workbench capabilities through bounded OEM and partner integrations
AI SECURITY SERVICES
Use focused reviews, deeper assessments, adversarial testing, hardening, and evidence work to answer a defined security or business decision. Each engagement uses the AI Security Workbench where it adds value, but the deliverable is the outcome and evidence—not access to an undifferentiated tool catalog.
Buyer situations
LAUNCH
Know what must be fixed, accepted, or evidenced before launch.
Decision answered
What must be fixed, accepted, or evidenced before this AI feature ships?
Typical duration
5–10 business days; first findings targeted within 5 business days.
Primary output
AI Launch Risk Memo and release-condition checklist
Best for
Teams preparing a selected AI feature, system, workflow, or release boundary for launch.
ASSESS
Map the AI product as a connected system, test material failure and abuse flows, qualify consequential paths, and turn the result into remediation and evidence.
Decision answered
What are the material security paths, control gaps, and remediation priorities across this AI product?
Typical duration
2–4 weeks, depending on scope.
Primary output
AI Product Security Assessment Report and prioritized remediation backlog
Best for
Teams that need a deeper product-security view than a bounded launch review.
ATTACK
Reproduce realistic adversarial behavior against the authorized AI system and preserve the evidence needed to understand, remediate, and retest what actually failed.
Decision answered
Which realistic abuse and adversarial behaviors can be reproduced, and what controls break them?
Typical duration
2–5 weeks, depending on scope.
Primary output
AI Red-Team Findings Register with reproduction and retest conditions
Best for
Teams that need authorized, realistic adversarial testing against a defined AI system boundary.
HARDEN
Map what the agent can actually cause to happen, reduce unnecessary authority, strengthen approvals and rollback, and verify the changed boundary.
Decision answered
Where can delegated authority, tool use, identities, permissions, approvals, or side effects exceed the intended boundary?
Typical duration
2–5 weeks, depending on scope.
Primary output
Agent Authority and Hardening Plan with verified boundary changes
Best for
Teams operating agents or AI workflows that can call tools, use credentials, or create consequential side effects.
VALIDATE
Determine what your guardrails and evals actually cover — and what still gets through.
Decision answered
What do the current guardrails and evals actually cover, and what failure cases still pass?
Typical duration
2–5 weeks, depending on scope.
Primary output
Guardrails and Evals Coverage Review with regression and release criteria
Best for
Teams with guardrails or evals already in place that need a coverage review before release.
RETRIEVAL
Test whether retrieval authorization, tenant boundaries, provenance, context integrity, and indirect-injection controls hold within the authorized system boundary.
Decision answered
Are retrieval authorization, tenant boundaries, provenance, context integrity, and indirect-injection controls holding?
Typical duration
2–4 weeks, depending on scope.
Primary output
RAG Security Findings and Retest Record
Best for
Teams shipping a RAG application, knowledge assistant, or retrieval-backed agent.
SELL
Turn existing security architecture, controls, testing, remediation, and evidence into defensible answers for customer and procurement review.
Decision answered
What AI security claims can the company safely make, and what evidence can support buyer review?
Typical duration
1–4 weeks, depending on scope.
Primary output
Approved buyer security narrative, answer bank, and evidence index
Best for
B2B AI companies and product teams repeatedly answering enterprise security and procurement questions.
BUILD
Define AI security ownership, intake, control expectations, release conditions, evidence requirements, exception handling, and an owned implementation backlog.
Decision answered
What ownership, controls, workflows, evidence, and backlog are required to operate AI security coherently?
Typical duration
4–10 weeks or retainer, depending on scope.
Primary output
AI Security Operating Model and owned implementation backlog
Best for
Organizations that need to turn fragmented AI policy and risk work into a security-operational program.
BASELINE
Establish a directional, analyst-reviewed baseline of AI security ownership, control coverage, evidence gaps, and priority work.
Decision answered
Where should the AI security program start, and which ownership, control, and evidence gaps need priority work?
Typical duration
1–3 weeks, depending on scope.
Primary output
Analyst-reviewed AI Security Program Scorecard baseline and prioritized roadmap
Best for
Leaders who need a bounded starting point before funding a larger assessment or program build.
Operating method
Map · Attack · Defend · Evidence
M.A.D.E. is the operating method across engagements, not the service taxonomy. The work maps the authorized system, tests material hypotheses, strengthens the boundary, and preserves the evidence needed for decisions and retest.
Scope boundary
Actual testing follows the authorized scope. Observations and retest results apply to the tested condition and boundary. An engagement does not establish certification, universal coverage, framework compliance, or the absence of vulnerabilities unless a separately defined and supportable determination explicitly says so.