Governance

Shadow AI: What Every CISO Missed in 2024

Unmanaged LLM usage is now the biggest unmapped data-exfiltration channel inside enterprises. Here is what the 2024 monitoring numbers show and what the pragmatic response looks like.

PromptShielder Security Team· Applied cryptography & privacy engineering March 11, 2025 6 min read

The gap between 'we banned it' and 'nobody uses it'

Every CISO survey in 2024 showed a majority claiming a ChatGPT policy exists. Every corresponding employee survey showed a majority using it anyway, at least weekly, and a large minority using it daily. The two data sets are describing the same organisation.

Why bans fail

  • Personal devices are outside your MDM.
  • Personal accounts are outside your SSO.
  • SaaS gateways only see traffic that goes through them.
  • Employees genuinely gain productivity — the incentive to route around the ban is strong.

What actually reduces exposure

Three moves compound. First, sanction a specific enterprise tier with a DPA — because ungrounded prohibition becomes ignored. Second, put a redaction control on the browser edge, on the personal account, on the device you do not own — because that is where the traffic actually lives. Third, publish a short list of concrete examples of what not to paste, with the redaction alternative for each. Policies that are three pages long get skimmed; policies that fit on one page get followed.

The 2024 numbers in one paragraph

Aggregating publicly disclosed browser-DLP telemetry from Cyberhaven, LayerX, Harmonic and Nightfall through 2024: 8–11% of monitored employees pasted material an internal review would have flagged; source code and customer records were the two most common categories; roughly three quarters of that traffic went to consumer LLM accounts not covered by a DPA; and the vast majority of it produced no incident ticket, no audit entry, and no legal-hold artefact.

Frequently asked

Should we block LLM domains at the firewall?+

It reduces measurable traffic and moves the rest to mobile networks. You lose visibility without losing usage. Do it only alongside a sanctioned alternative.

Do browser DLP tools solve this?+

They observe. They do not always redact before send. A redaction-on-paste control complements them.

Sources
GovernancePrivacy
Try PromptShielder

Redact prompts before they leave your browser.

Free to try. Nothing is sent to a server. Ever.