Guilty Until Proven Durable: AI Casts A Long (Security) Shadow

研报英文原文证据摘录

Guilty Until Proven Durable: AI Casts A Long (Security) Shadow

Shadow AI = Shadow IT On Steroids

The Shadow AI market is still early, but we believe the control stack required to support GenAI

SaaS apps and autonomous agents will need to scale quickly as these tools gain access to

enterprise data, applications, and workflows. The Samsung ChatGPT source-code leak in March

2023 remains the most cited example: employees reportedly pasted confidential code and

internal notes into ChatGPT, creating a clear enterprise data-loss event. A more recent — and

more pointed — example is Anthropic’s Fable 5. Launched in June 2026, the model was briefly

pulled under a US export-control ban over its cyber capabilities before being reinstated in July

with a hardened safety classifier, underscoring how quickly frontier capability is outrunning the

controls meant to contain it.

We believe Shadow AI could become materially larger than Shadow IT because it builds on

existing unsanctioned-app risk while adding new vectors around model access, data leakage,

agent autonomy, prompt injection, governance, and observability. Demand should be driven by

three forces: rising regulatory pressure, expanding security and governance budgets, and the

need for new, Shadow AI-specific controls as enterprises deploy GenAI and LLMs — a

combination that opens up an incremental TAM opportunity.

Due to the speed at which threat vectors are evolving, regulation is building quickly on both sides

of the Atlantic. In the US, a June 2026 executive order established an early-access review

process for frontier models that materially advance national-security capabilities; FINRA’s

Regulatory Notice 24-09 already applies existing securities rules to GenAI and LLM use; and the