~/f4n6 $ grep -r "Shadow AI usage exceeds reported rates by 30x in DNSFilter report" ./investigations/ --include="*.md"

Shadow AI usage exceeds reported rates by 30x in DNSFilter report

Jeff Davies 05 Nov 2026 5 min read

1. Executive summary

DNSFilter's The Cost of Shadow AI report, covered by Help Net Security on 2026-11-05, states that 46.1% of surveyed organisations had an employee connect an unapproved AI tool to their Google Workspace, Microsoft 365, or SSO environment, and that 20.1% said such a connection caused a security incident. DNSFilter's own network telemetry for August 2026, drawn from 3.4 billion generative AI requests across more than 43,000 North American organisations, found the typical organisation connected to five distinct AI vendors and 15.2% to more than 10, roughly 30 times the rate IT professionals self-reported. Only 0.75% of those requests were blocked. The report is a survey and telemetry study, not a report of a specific intrusion: no CVE, threat actor, malware or victim is named, and no exploitation is recorded. For EMEA financial services the exposure is third-party and identity risk from unsanctioned OAuth grants into productivity and SSO estates, which sits inside existing ICT third-party and incident-reporting duties rather than presenting a new technical vulnerability to patch.

2. Regulatory framing

No specific DORA/NIS2 article is directly engaged by this item.

3. Technical analysis & attack chain

The item is a survey and telemetry report. It describes no intrusion, names no victim, no malware and no threat actor, and gives no initial-access vector, payload, persistence mechanism, command-and-control channel or exfiltration path. There is no attack chain to reconstruct from confirmed steps. What follows is what the report establishes about exposure and what it does not.

The exposure the report measures is unsanctioned OAuth and SSO authorisation. Employees connect third-party AI tools to Google Workspace, Microsoft 365 or an SSO environment. In organisations where employees use AI tools for work, 36.8% allow employees to grant that access and notify IT afterwards, and 6.5% require no notification at all. The report does not describe the OAuth scopes requested, the token lifetimes, the vendors involved, or how a connection translates into data access. It records only that 46.1% of respondents had at least one such connection and 20.1% said a connection caused a security incident. The nature of those incidents is not stated in public reporting.

The oversight gap is measured on the security-tool side. 61.5% of respondents cannot say that every security tool in their environment is properly configured, and 61.8% cannot say that every tool is actively monitored. 54% worry the organisation has significant AI costs that procurement, software asset management and security monitoring teams cannot track, rising to 58.8% for IT managers. The report states that budget approvers are the least concerned about untracked spending while practitioners know AI tools have been purchased but cannot see the details.

The telemetry comparison is the report's central claim. Asked how many AI tools employees use for work, 55.3% of respondents said three to five and 0.5% said more than 10. DNSFilter's August 2026 measurement across two business days found the typical organisation connected to five distinct AI vendors and 15.2% connected to more than 10, which the report describes as approximately 30 times the rate reported by IT professionals. Only 0.75% of generative AI requests were blocked. The report notes that allowing AI can be a deliberate policy choice, so the block rate alone does not establish a control failure.

Confidence caveats. The survey figures are self-reported by respondents and the telemetry is DNSFilter's own, so the 30x comparison rests on a single vendor's data set covering North American organisations; it is single-sourced and has not been independently corroborated. The report does not state the survey sample size, the survey period, the respondent sectors, or whether any respondent is a financial services firm. The 20.1% incident figure is a respondent claim with no incident detail attached. Adverse Trace assesses that the pattern described, employees granting OAuth access to unapproved AI tools before or without IT notification, is consistent with the third-party and identity exposure EMEA financial clients already carry, but the report provides no evidence of a specific compromise at a specific institution.

4. Mitigation & containment

The report names no product version, patch, command, file path or registry key, so no technical remediation can be lifted from it. The steps below target the process and identity controls the report's findings implicate. Where a control depends on a platform capability, confirm the exact setting in your own tenant before enforcing it.

P1 within 24h

  • Pull the OAuth grant inventory for Google Workspace, Microsoft 365 and your SSO provider and list every third-party application with an active grant, its granting user, the scopes held and the grant date. The report's 46.1% figure rests on connections of exactly this kind being visible after the fact.
  • For any AI tool on that list that is not on your approved register, revoke the grant and notify the granting user's line manager. Record the revocation.
  • Check whether your SSO or identity provider is configured to require admin consent for new third-party application grants. Where it is not, the 36.8% "grant then notify" and 6.5% "no notification" patterns in the report apply directly to your estate.

P2 within 72h

  • Reconcile the AI tools your IT and security teams believe are in use against the vendors seen in proxy, CASB or DNS logs. The report's gap between 0.5% self-reporting more than 10 tools and 15.2% actually connecting to more than 10 is the measurement to reproduce internally.
  • Confirm that every security tool in the estate is configured as intended and is actively forwarding logs, the two areas where 61.5% and 61.8% of respondents respectively could not give assurance.
  • Bring procurement, software asset management and security monitoring onto one list of AI tool spend and AI tool connections. The report's 54% untracked-cost figure is the symptom of those teams working from separate records.

P3 within 7 days

  • Publish an AI tool approval process that names who may approve a connection to corporate data and requires that approval before the grant, not after.
  • Set a blocking or alerting policy for generative AI categories in the secure web gateway or DNS layer, and record whether each category is deliberately allowed. The report's 0.75% block rate is only meaningful against a stated policy.
  • Add OAuth grant creation for new third-party applications to the SIEM alert set and to the incident classification process, so a connection that later causes an incident is captured as an ICT-related event.

5. Indicators of compromise

No indicators of compromise available in the source material.

The report describes no malicious file, domain, IP address, hash or command. Its observations are statistical and behavioural, and the behaviours it reports (OAuth grants to unapproved AI tools, unmonitored security tooling, untracked AI spend) are organisational patterns rather than artefacts attributable to a threat. There is no atomic indicator set to publish and no copyable block.

6. Detection

Insufficient indicators to author detection rules.

The report gives no file strings, command-line flags, mutex names, scheduled-task or service names, registry keys or network artefacts belonging to a threat. Any rule built from its figures or vendor names would detect reporting about shadow AI, not shadow AI itself.

7. Sources

  • Help Net Security, Companies are spending more on AI governance and getting less control, https://www.helpnetsecurity.com/2026/11/05/dnsfilter-shadow-ai-cost-report/, 2026-11-05. Reports DNSFilter's The Cost of Shadow AI survey and DNSFilter's August 2026 network telemetry.

8. Adverse Trace position

Adverse Trace assesses this as an informational item with low direct severity for EMEA financial services. It reports no vulnerability, no exploited CVE, no named actor and no victim, and the CISA KEV catalogue and NVD carry no entry for it. The client impact is indirect: unsanctioned OAuth grants into Google Workspace, Microsoft 365 and SSO estates are a live third-party and identity exposure, and the report's own figures (46.1% of respondents with at least one such connection, 20.1% reporting an incident from one) show the exposure is common. The 30x discrepancy between self-reported and observed AI tool use rests on a single vendor's North American telemetry and should be treated as indicative, not as a measurement of any client's estate. We will keep the OAuth grant inventory and AI tool approval process on the client checklist for third-party reviews, and we will revisit the assessment if DNSFilter publishes the underlying methodology or if a named incident is linked to an unapproved AI tool connection. Clients wanting to test their own exposure against the report's method should start with the P1 grant inventory above.


Read the original source →

Published via PulseTrace — Adverse Trace threat intelligence.

Post this to LinkedIn
Formatting is converted automatically — headings, bullets, a link back & hashtags. Paste straight in.
J
Jeff Davies