Enterprise Security in the Agentic AI Era

Shadow AI

3 min read

Shadow AI

Shadow AI is the use of unauthorized AI tools in the workplace: the AI-era counterpart of Shadow IT. The same dynamic (unsanctioned tooling adopted by individuals to solve real problems) but with a higher data-leakage blast radius because every interaction puts text or files into a third-party AI system.

Scale

The 75% / 78% figures are typically cited via Knostic, but the primary source is Microsoft’s Work Trend Index 2025. The wiki should cite Microsoft directly rather than the downstream vendor. Independently corroborated:

  • Stanford HAI AI Index 2025: 78% business AI adoption (up from 55% in 2023); methodology disclosed; cross-year comparable
  • McKinsey State of AI (Nov 2025): 88% of orgs use AI in ≥1 function; only 6% are “high performers” (n=1,993, 105 countries)
  • BCG “Build for the Future 2025”: 72% regular use; only 13% have AI agents in production vs 56% experimenting; the agentic-vs-GenAI gap is load-bearing
  • IBM Cost of a Data Breach 2025: 20% of orgs experienced shadow-AI breaches; cost $670K above average

Directional agreement: rate is high enough that “block all unsanctioned AI” is impractical. The agentic deployment lag (BCG’s 13% in production) is the wiki’s CMM “L1 Initial” anchor.

See Source Triangulation Audit 2026-05-02 §Claim 2 for full triangulation.

Canonical Incident

The Samsung incident (early 2023): engineers pasted proprietary chip-design code into ChatGPT for debugging help. The data left the corporate boundary and (per OpenAI’s then-policy) became eligible for model improvement. Samsung subsequently restricted external GenAI use enterprise-wide. This incident is now the canonical Shadow AI cautionary tale and is referenced across vendor and academic materials.

Risk Profile

Shadow AI extends standard Shadow IT risks plus:

RiskShadow ITShadow AI
Unauthorized vendor relationship✓✓
Unaccounted data egress✓✓ + likely enriched (extracted, summarized, transformed)
Compliance gap (GDPR, HIPAA, etc.)✓✓ + harder to remediate (model training is irreversible)
Security review bypass✓✓ + AI-specific issues (prompt injection, hallucination, vector poisoning)
Vendor lock-in / continuity risk✓✓
Inference exposure (Inference Exposure (and Retrieval Exposure))✗✓ (unique to AI)
Training-data contamination✗✓ (corporate IP enters base models)
Cross-customer leakage✗✓ (early incidents, e.g. ChatGPT outage 2023, showed cross-session exposure)

Mitigation

The Knostic article frames mitigation as a triad:

  1. Governance policies: sanctioned AI list with clear escalation path for adding new tools
  2. Usage monitoring: DLP / network telemetry / browser-extension visibility into AI tool usage
  3. Employee training: what is and is not safe to put into an AI tool, with concrete examples

Mature implementations add:

  • Sanctioned alternatives provided proactively. If users have a sanctioned tool that does what they need, BYOAI rates drop.
  • Discovery via AI-SPM. Inventory production-grade AI assets, including locally installed MCP servers and IDE extensions.
  • Knowledge-layer controls (Oversharing Controls for AI Search) to limit damage when sanctioned AI is misused for unsanctioned data.

Distinction from Sanctioned AI

The bright-line rule: sanctioned ≠ safe. Microsoft Copilot, Glean, and Gemini are all sanctioned in many enterprises and still drive significant oversharing risk. Shadow AI is the unsanctioned-tool problem; oversharing is the sanctioned-tool problem. Both apply.

See Also

Sources

Graph View

Backlinks