Audience: CISOs, AI governance leads, risk managers, compliance officers, board advisors

Critical Assessor Notice: Boundaries of Cryptographic Evidence

SWT3 witness anchors prove that specific operational controls were active at a specific point in time. They do not replace the assessor's independent judgment, professional expertise, or regulatory authority. Assessors must verify that anchored evidence is sufficient, appropriate, and relevant to the specific assessment context. IBM's breach data quantifies the cost of governance gaps; SWT3 provides the evidence layer, not the governance decisions themselves.

1. Executive Summary

IBM's 2026 Cost of a Data Breach Report surveyed 602 organizations across 16 countries and 17 industries. The central finding: AI governance failure, not AI itself, is driving breach costs.

$4.99M Global average breach cost
(12% increase, 21-year high)
68% Of breached organizations had
no AI governance policy
43% Experienced shadow AI incidents
(doubled from 20%)
92% Of AI-affected breaches lacked
proper access controls

The report measures six governance controls. Five of the six declined year over year. Only 19% of organizations reported coordination between governance and security teams. Shadow AI incidents more than doubled, and AI-enabled attacks now account for one in four malicious breaches.

This guide maps IBM's findings to specific SWT3 procedures. For each gap IBM identified, there is a corresponding procedure that produces verifiable, timestamped, cryptographic evidence. The goal is not to solve every problem the report raises. It is to show what an evidence-based response looks like for each one.

2. The Six Governance Controls IBM Measured

IBM measured adoption rates for six AI governance controls across breached organizations. Five declined year over year. The only control that gained adoption was adversarial testing, and even that reached only 25%.

Governance Control 2026 2025 Trend SWT3 Procedure
Strict AI deployment approvals 38% 45% -7 AI-GOV.1, AI-HITL.1
AI governance technology 33% 39% -6 Witness infrastructure
Governance frameworks 33% 39% -6 36 framework crosswalks
Employee AI risk training 30% 36% -6 241 compliance guides
Audits for unauthorized AI 29% 34% -5 AI-DENSITY.1, AI-AUDIT.1
Red teaming / adversarial testing 25% 22% +3 AI-GRD.1, AI-GRD.2

What each mapping means

AI-GOV.1 + AI-HITL.1

Strict AI deployment approvals

IBM found that fewer organizations required formal approval before deploying AI in 2026 than in 2025. The gap is not policy. It is proof. AI-GOV.1 records that a governance review occurred, who participated, and what the outcome was. AI-HITL.1 records that a human reviewed and approved the deployment. Both produce timestamped, cryptographic anchors that exist before the system goes live. The pre-inference gate (authorization_id) blocks inference until an approved authorization exists.

AI-DENSITY.1 + AI-AUDIT.1

Audits for unauthorized AI

Only 29% of breached organizations audited for unauthorized AI use. AI-DENSITY.1 records the ratio of witnessed events to total events over a time window. When coverage drops from 100% to 2%, the density anchor records the gap before an auditor arrives. AI-AUDIT.1 records the integrity of the audit trail itself, producing a hash of the evidence chain so tampering is detectable. Together they answer two questions: "Is everything being witnessed?" and "Has the evidence been altered?"

AI-GRD.1 + AI-GRD.2

Red teaming and adversarial testing

This was the only control that gained adoption, reaching 25%. AI-GRD.1 records input-side guardrail activation (what was blocked before it reached the model). AI-GRD.2 records output-side filter results (what the model produced and whether it passed content classification). Both produce evidence that guardrails were tested and active, not just configured.

3. The Shadow AI Gap

IBM finding: Shadow AI incidents more than doubled to 43% of breached organizations, up from 20%. Average breach cost with shadow AI involvement: $5.39 million. Outcomes included data loss (49%), operational disruption (42%), reputational damage (35%), and regulatory fines (~20%).
IBM 2026 Cost of a Data Breach Report

Shadow AI is any AI system operating without organizational visibility or governance. The cost difference is clear: $5.39 million with shadow AI versus $4.99 million overall. The gap is not about banning AI use. It is about knowing what is running.

How the evidence layer addresses this

AI-ID.1

Agent identity

Every AI system that connects to the witness infrastructure receives a registered identity. The agent_id field appears on every anchor. If an AI system is producing inferences without an agent_id, it is shadow AI by definition. The presence or absence of identity registration is the first diagnostic.

AI-INF.1

Inference witnessing

Every inference call produces an anchor with the model, the clearing level, and a timestamp. AI systems that produce inferences without witness anchors are invisible to governance. Wrapping inference calls with the witness SDK is a single line of code (witness.wrap(client)). From that point forward, every inference is recorded independently of the AI system itself.

AI-DENSITY.1

Anchor density monitoring

Even after wrapping AI systems, coverage can degrade over time. A system update, a configuration change, or a new deployment path can silently break witnessing. AI-DENSITY.1 records the ratio of witnessed events to total events. When coverage drops, the density anchor captures the gap. This is how you detect shadow AI that starts as governed AI and drifts into ungoverned territory.

The pattern: AI-ID.1 registers the system. AI-INF.1 witnesses every inference. AI-DENSITY.1 catches coverage gaps. If all three are active, shadow AI produces a detectable signal: either missing identity, missing inference anchors, or declining density ratios.

4. The Access Control Gap

IBM finding: 92% of organizations that experienced an AI-related breach lacked proper access controls at the time of the incident. Only 40% of all breached organizations applied access controls to AI systems.
IBM 2026 Cost of a Data Breach Report

This is the largest gap IBM measured. Access controls for traditional IT systems are well understood. Access controls for AI systems are not. The question is not just "who can use the AI system" but "what is the AI system allowed to do, access, and decide."

AI-ACC.1

Access control witnessing

Records every access attempt: who requested access, what scope they requested, whether access was granted, and whether the request fell within the declared scope. Each access event produces an anchor. The absence of access anchors for a data source the AI system used is evidence that access controls were bypassed or missing.

AI-DEL.1 + AI-DEL.2

Delegation tree and delegation boundary

AI-DEL.1 records what the AI system is permitted to do: which tools, which data, which decisions. AI-DEL.2 records what it is not permitted to do: blocked tools, restricted scopes, actions that require escalation. Together they define the operating envelope. If the AI system acts outside its declared boundary, either the gatekeeper blocks it or the action produces an anchor that proves the boundary was crossed.

AI-TOOL.1

Tool call witnessing

Every tool call an AI agent makes produces an anchor: which tool, how long it took, whether it succeeded. For agentic AI systems that can call external APIs, execute code, or access databases, this is the access control audit trail. You can see exactly which tools were called, when, and whether they fell within the delegation boundary.

Clearing Levels 0-3

Data classification enforcement

Four clearing levels control what data flows through the witness infrastructure. Level 0 (analytics) strips all identifying content. Level 3 (classified) retains full context. The clearing level is set per deployment and enforced by the SDK. This maps directly to IBM's access control gap: instead of asking whether access controls exist, the clearing level system enforces them at the data layer.

5. The Cross-Functional Coordination Gap

IBM finding: Only 19% of breached organizations reported coordination between governance teams and security teams.
IBM 2026 Cost of a Data Breach Report

Governance teams write policies. Security teams detect threats. When these teams do not share a common evidence layer, policies cannot reference real operational data, and security alerts cannot trigger governance responses. The gap is not communication. It is shared infrastructure.

AI-CHAIN.1

Chain-of-custody handoffs

When an AI system moves between teams (from development to security review to production deployment), AI-CHAIN.1 records each handoff. The anchor captures who handed off, who received, and what state the system was in. Both governance and security teams can query the same chain of custody without relying on email threads or meeting notes.

AI-TRUST.1 + AI-TRUST.2

Cross-system trust verification

In multi-agent architectures, AI-TRUST.1 verifies another system's compliance state. AI-TRUST.2 presents credentials. These procedures create a trust mesh where systems can verify each other's governance status cryptographically. When governance and security tools share the same trust mesh, their evidence is interoperable by default.

Anchor Verifier

Shared verification endpoint

The public anchor verifier at /verify lets anyone in the organization paste an anchor fingerprint and confirm it is valid. Governance teams can verify security evidence. Security teams can verify governance approvals. Neither team needs access to the other's tools. The verification is cryptographic and independent of the system that produced the evidence.

The pattern: When governance and security share the same evidence substrate (witness anchors), coordination becomes a data problem instead of an organizational problem. Both teams query the same ledger. Verification is self-service. The 19% coordination rate IBM measured reflects a missing infrastructure layer, not a missing process.

6. AI Attack Vectors and Their Cost

IBM measured the average breach cost for six categories of AI-related security incidents. One in four malicious attacks in the study were AI-driven, a 56% increase year over year. The average cost of an AI-driven attack was $6.04 million, roughly $1 million more than non-AI attacks.

Attack Vector Avg. Cost Evidence Procedure What the Evidence Shows
Model inversion $6.07M AI-MDL.5, AI-PROV.1 Which model version was deployed, who trained it, and whether weight integrity was maintained
Prompt injection $5.89M AI-GRD.1, AI-GRD.2 Whether input guardrails were active and whether output filters caught the injected content
Cloud misconfiguration (AI workloads) $5.25M AI-HW.1, AI-MCP.1 Hardware attestation state and MCP connection security posture at time of incident
Malicious models $4.94M AI-PROV.1, AI-MDL.5/6/7 Model provenance chain: base model, training pipeline, every fine-tune and quantization step
Model evasion $4.72M AI-DRIFT.2 Consequence-mapped drift thresholds: whether model behavior crossed a defined boundary
Compromised apps, APIs, plug-ins $4.37M AI-TOOL.1, AI-ACC.1 Which tools and APIs were called, whether access was authorized, and whether payloads were signed
62% of AI-driven attacks targeted critical infrastructure sectors. The cost table above applies disproportionately to energy, water, telecom, transportation, and financial services. Organizations in these sectors face both higher likelihood and higher cost of AI-related breaches.

7. The Cost of Speed vs. the Cost of Evidence

IBM finding: Breaches that took longer than 200 days to identify and contain cost an average of $5.65 million. Breaches resolved faster cost $4.32 million. The mean time to identify and contain a breach was 247 days, up from 241 days the prior year.
IBM 2026 Cost of a Data Breach Report

The cost difference between fast and slow resolution is $1.33 million. Detection speed depends on two things: knowing something happened, and having enough context to understand what happened. Continuous witnessing addresses both.

How continuous evidence reduces detection time

Every anchor includes a millisecond-precision timestamp. When an incident occurs, the anchor chain provides a timeline of exactly what the AI system did, when it did it, and what authorization it was operating under. Investigators do not start from zero. They start from a cryptographically verified timeline.

AI-EMRG.1 records the emergency response lifecycle. The lifecycle chain links the detection, the authorization to restrict, the restriction itself, and the resumption into a single chain of evidence connected by a shared identifier. The time between each stage is measurable. The organization can define acceptable response windows in advance and measure whether they were met.

AI-INCIDENT.1 records incident events as they happen. Instead of reconstructing an incident timeline weeks later from logs, interview notes, and memory, the incident witness captures events as they occur. The evidence exists in real time, not after the fact.

The cost-reduction factors IBM identified align with the evidence layer. IBM's top cost-reducing factors were DevSecOps ($253,805 saved), identity and access management ($225,622), key lifecycle management ($214,923), and encryption ($213,478). SWT3 anchors are cryptographically signed, identity-linked, access-controlled, and integrated into DevSecOps pipelines through the GitHub Action and CI/CD gate. These are not separate investments. They are properties of the evidence layer itself.

8. Your Next Board Conversation

IBM's report will reach your board. These are the questions they will ask and the answers the evidence layer provides.

"Do we have an AI governance policy?"

68% of breached organizations did not. With SWT3, the question shifts from "do we have a policy document" to "do we have evidence that governance decisions are being made and enforced." Every AI-GOV.1 anchor is proof that a governance review occurred. Every AI-HITL.1 anchor is proof that a human approved a deployment. The anchor chain is the policy made operational.

"How do we know we do not have shadow AI?"

43% of breached organizations discovered shadow AI during the breach. AI-DENSITY.1 reports the ratio of witnessed to total AI events continuously. If the ratio drops, something is running unwitnessed. This is a leading indicator, not a finding that arrives with the breach notification.

"Are our AI systems access-controlled?"

92% of AI-affected breaches lacked access controls. AI-ACC.1 records every access attempt. AI-DEL.2 records what each system is not permitted to do. The anchor chain shows, for any given day, exactly which AI systems had access to which data and tools, and whether that access was within the declared scope.

"How quickly would we detect an AI-related incident?"

IBM's mean time to identify and contain a breach was 247 days. Continuous witnessing creates a real-time evidence trail. When an anomaly appears in the anchor chain (a drift threshold crossed, a density ratio dropping, an access attempt outside scope), it is detectable in the current data, not in a quarterly audit.

"What would this cost us?"

IBM measured the U.S. average at $11.5 million per breach, 2.3 times the global average. AI-driven attacks averaged $6.04 million. Shadow AI breaches averaged $5.39 million. The cost of evidence infrastructure is a fraction of the cost of a single incident.

9. Getting Started

The IBM report measures gaps. SWT3 produces evidence. The shortest path from reading this guide to producing your first verifiable governance artifact is three steps.

Step 1: Install the SDK

pip install swt3-ai          # Python
npm install @tenova/swt3-ai   # TypeScript

Step 2: Witness your first inference

# Python
from swt3_ai import Witness
from openai import OpenAI

witness = Witness(
    endpoint="https://sovereign.tenova.io",
    api_key="axm_...",
    tenant_id="YOUR_TENANT",
)

client = witness.wrap(OpenAI())
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
)

receipts = witness.flush()

Step 3: Verify the anchor

Open the Anchor Verifier, paste the fingerprint from your receipt, and confirm it verifies. That fingerprint is your first piece of cryptographic governance evidence.

For a zero-configuration demo that requires no API keys or accounts:

python -m swt3_ai.demo    # Python
npx swt3-demo             # TypeScript
Further reading: SDK Documentation | Assessment Playbook | Quickstart Guide | Create a Free Account