Verax · AI Governance Assessment

AI Usage & Vendor Risk Assessment

Prepared for Acme Corp · Generated 2026-07-22 16:25 UTC

Executive summary

AI tools in use

AI vendors detected

151

events to vendors that train on inputs

users involved

268

total findings

Severity	Findings
critical	150
medium	94
low	24

Executive impact

60.0h

Analyst time saved (10 vendors x 6.0h @ $150/h)

$9,000

Assessment labor value delivered

high-risk standing-access items (7 EU AI Act-flagged)

$960

est. annual tool consolidation (4 redundant tool(s) @ $240/seat/yr)

Industry average breach cost (IBM Cost of a Data Breach 2024) — benchmark for the standing-access attack surface, not a guaranteed loss. Benchmark breach cost: $4,880,000. EU AI Act penalties reach 7% of global annual turnover for the most serious violations.

Vendor risk assessments

Perplexity AI critical

Observed activity 30 events · users: asmith, bwayne, ckent, ckent@acme.com, dprince, eturing, jdoe, jdoe@acme.com

Trains on customer input consumer_tier

Data retention May use data to improve services unless AI data retention disabled

Data region US

Certifications SOC 2 Type II

DPA available yes

Breach history none known

EU AI Act GPAI-based service

Risk drivers OAuth grant to AI app 'Perplexity': openid profile email · Unsanctioned tool (not on client allowlist) · Vendor base risk: high · Vendor trains on inputs (consumer_tier)

Recommended remediation

FortiGate: Block Perplexity AI at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-Perplexity-AI"
        config entries
            edit 1
                set url "perplexity.ai"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Microsoft Graph: Revoke OAuth grant to Perplexity

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp5-3"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0006 --set accountEnabled=false

Microsoft Graph: Revoke OAuth grant to Perplexity

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp5-0"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0006 --set accountEnabled=false

OpenAI ChatGPT critical

Observed activity 29 events · users: asmith, asmith@acme.com, bwayne, ckent, ckent@acme.com, dprince, eturing, jdoe

Trains on customer input consumer_tier

Data retention 30 days (API); chat history used for training on Free/Plus unless opted out

Data region US

Certifications SOC 2 Type II

DPA available yes

Breach history 2023 Redis bug exposed some chat titles

EU AI Act GPAI provider; transparency obligations apply

Risk drivers OAuth grant to AI app 'ChatGPT': openid profile email Mail.Read Files.Read.All offline_access · Sensitive scopes: Files.Read.All, Mail.Read · Unsanctioned tool (not on client allowlist) · Vendor base risk: high · Vendor trains on inputs (consumer_tier)

Recommended remediation

FortiGate: Block OpenAI ChatGPT at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-OpenAI-ChatGPT"
        config entries
            edit 1
                set url "chatgpt.com"
                set type simple
                set action block
            next
            edit 2
                set url "openai.com"
                set type simple
                set action block
            next
            edit 3
                set url "chat.openai.com"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Microsoft Graph: Revoke OAuth grant to ChatGPT

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp0-1"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0001 --set accountEnabled=false

Microsoft Graph: Revoke OAuth grant to ChatGPT

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp0-3"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0001 --set accountEnabled=false

Google Gemini critical

Observed activity 25 events · users: bwayne, ckent, dprince, dprince@acme.com, eturing, jdoe, jdoe@acme.com

Trains on customer input consumer_tier

Data retention Consumer activity reviewed/used to improve services unless disabled

Data region Global

Certifications ISO 27001, SOC 2 Type II

DPA available yes

Breach history none known

EU AI Act GPAI provider; transparency obligations apply

Intelligence verified 2026-07-18 · sources: [1]

Risk drivers OAuth grant to AI app 'Gemini': openid profile Calendars.Read Contacts.Read · Sensitive scopes: Calendars.Read, Contacts.Read · Unsanctioned tool (not on client allowlist) · Vendor base risk: high · Vendor trains on inputs (consumer_tier)

Recommended remediation

FortiGate: Block Google Gemini at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-Google-Gemini"
        config entries
            edit 1
                set url "gemini.google.com"
                set type simple
                set action block
            next
            edit 2
                set url "bard.google.com"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Microsoft Graph: Revoke OAuth grant to Gemini

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp2-4"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0003 --set accountEnabled=false

Microsoft Graph: Revoke OAuth grant to Gemini

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp2-0"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0003 --set accountEnabled=false

GitHub Copilot critical

Observed activity 23 events · users: asmith, bwayne, ckent, dprince, eturing, jdoe

Trains on customer input consumer_tier

Data retention Business/Enterprise: prompts not retained/used for training; Individual differs

Data region US

Certifications SOC 2 Type II

DPA available yes

Breach history none known

EU AI Act Code suggestions; IP/licensing considerations

Risk drivers Unsanctioned tool (not on client allowlist) · Vendor base risk: medium · Vendor trains on inputs (consumer_tier)

Recommended remediation

FortiGate: Block GitHub Copilot at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-GitHub-Copilot"
        config entries
            edit 1
                set url "copilot.github.com"
                set type simple
                set action block
            next
            edit 2
                set url "api.githubcopilot.com"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Grammarly critical

Observed activity 23 events · users: asmith, bwayne, ckent, dprince, eturing, jdoe

Trains on customer input consumer_tier

Data retention Processes text typed across apps; enterprise has stricter controls

Data region US

Certifications SOC 2 Type II, ISO 27001

DPA available yes

Breach history 2018 browser extension token bug

EU AI Act Text processing across endpoints

Risk drivers Unsanctioned tool (not on client allowlist) · Vendor base risk: medium · Vendor trains on inputs (consumer_tier)

Recommended remediation

FortiGate: Block Grammarly at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-Grammarly"
        config entries
            edit 1
                set url "grammarly.com"
                set type simple
                set action block
            next
            edit 2
                set url "app.grammarly.com"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Microsoft Copilot critical

Observed activity 18 events · users: asmith, bwayne, ckent, dprince, eturing

Trains on customer input consumer_tier

Data retention Enterprise Data Protection for entra-signed-in commercial users; consumer differs

Data region Global

Certifications ISO 27001, SOC 2 Type II

DPA available yes

Breach history none known

EU AI Act GPAI provider; transparency obligations apply

Risk drivers Unsanctioned tool (not on client allowlist) · Vendor base risk: medium · Vendor trains on inputs (consumer_tier)

Recommended remediation

FortiGate: Block Microsoft Copilot at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-Microsoft-Copilot"
        config entries
            edit 1
                set url "copilot.microsoft.com"
                set type simple
                set action block
            next
            edit 2
                set url "copilot.cloud.microsoft"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Otter.ai critical

Observed activity 3 events · users: asmith@acme.com, dprince@acme.com, jdoe@acme.com

Trains on customer input consumer_tier

Data retention Audio and transcripts stored; may be used to improve speech recognition

Data region US

Certifications SOC 2 Type II

DPA available yes

Breach history none known

EU AI Act AI-generated transcripts; GDPR obligations for EU personal data in audio

Risk drivers OAuth grant to AI app 'Otter.ai': openid profile Calendars.Read · Sensitive scopes: Calendars.Read · Vendor trains on inputs (consumer_tier)

Recommended remediation

Microsoft Graph: Revoke OAuth grant to Otter.ai

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp4-1"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0005 --set accountEnabled=false

Microsoft Graph: Revoke OAuth grant to Otter.ai

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp4-0"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0005 --set accountEnabled=false

Microsoft Graph: Revoke OAuth grant to Otter.ai

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp4-4"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0005 --set accountEnabled=false

Fireflies.ai critical

Observed activity 1 events · users: tenant-wide

Trains on customer input no

Data retention Zero Data Retention policy for meeting content (audio, video, transcripts, summaries) with all third-party vendors/subprocessors — not stored, accessed, or used for AI model training after processing. Customer retains ownership; deleting a meeting purges it immediately and irreversibly. General account data retained ~12 months by default, user-controlled.

Data region US

Certifications SOC 2 Type II, GDPR, HIPAA

DPA available yes

Breach history none known

EU AI Act Contractual zero-retention and no-training-on-content commitments plus SOC 2/GDPR/HIPAA certification give a strong compliance posture for EU AI Act meeting-transcription use cases.

Intelligence verified 2026-06-27 · sources: [1] [2] [3] [4]

Risk drivers Matches the standing-OAuth-access pattern exploited in the 2025 Salesloft Drift breach (700+ orgs incl. Zscaler, Palo Alto). · OAuth grant to AI app 'Fireflies.ai Notetaker': openid Calendars.Read Chat.Read offline_access · Sensitive scopes: Calendars.Read, Chat.Read · Tenant-wide admin consent (all users)

Recommended remediation

Microsoft Graph: Revoke OAuth grant to Fireflies.ai Notetaker

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp3-all"

Revokes the tenant-wide admin consent for this app. To disable the app entirely: az ad sp update --id app-0004 --set accountEnabled=false

Slack AI medium

Observed activity 92 events · users: asmith, bwayne, ckent, dprince, eturing, jdoe

Trains on customer input no

Data retention Slack does not use Customer Data (messages, files) to train generative AI/LLMs without affirmative opt-in consent; uses RAG so LLM providers never retain data; separately, legacy non-generative ML features (e.g., emoji/channel recommendations) analyze customer data by default with an opt-out available via support request.

Data region US

Certifications SOC 2 Type II, ISO 27001

DPA available yes

Breach history none known for Slack AI specifically (Slack faced 2015 credential breach of user profile data, unrelated to AI features)

EU AI Act Embedded GenAI assistant inside core collaboration product; limited-risk transparency obligations apply; 2024 public controversy over ambiguous ToS language prompted clarified policy in April 2025.

Intelligence verified 2026-06-27 · sources: [1] [2] [3]

Risk drivers Unsanctioned tool (not on client allowlist) · Vendor base risk: low

Recommended remediation

FortiGate: Block Slack AI at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-Slack-AI"
        config entries
            edit 1
                set url "slack.com"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Anthropic Claude low

Observed activity 24 events · users: asmith, bwayne, ckent, dprince, dprince@acme.com, eturing, eturing@acme.com, jdoe, jdoe@acme.com

Trains on customer input no

Data retention Does not train on inputs by default; retention per plan

Data region US

Certifications SOC 2 Type II, ISO 27001

DPA available yes

Breach history none known

EU AI Act GPAI provider; transparency obligations apply

Risk drivers OAuth grant to AI app 'Claude': openid profile email User.Read · Vendor base risk: low

Recommended remediation

FortiGate: Block Anthropic Claude at the firewall

config webfilter urlfilter
    edit 0
        set name "Verax-Block-Anthropic-Claude"
        config entries
            edit 1
                set url "claude.ai"
                set type simple
                set action block
            next
            edit 2
                set url "anthropic.com"
                set type simple
                set action block
            next
        end
    next
end

Attach this urlfilter to a webfilter profile applied to your outbound policy.

Microsoft Graph: Revoke OAuth grant to Claude

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp1-0"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0002 --set accountEnabled=false

Microsoft Graph: Revoke OAuth grant to Claude

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp1-5"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0002 --set accountEnabled=false

Microsoft Graph: Revoke OAuth grant to Claude

az rest --method DELETE --uri "https://graph.microsoft.com/v1.0/oauth2PermissionGrants/grant-sp1-4"

Revokes the user consent for this app. To disable the app entirely: az ad sp update --id app-0002 --set accountEnabled=false

Compliance exposure

Framework	Findings flagged
EU AI Act	151
GDPR	151
ISO 42001	151

Methodology. This assessment is produced agentlessly from existing network telemetry (firewall / secure web gateway logs) matched against the Verax AI vendor intelligence base. Findings are metadata-based (who used which AI service, when, how much); prompt or file content is never inspected. Vendor data-handling facts are sourced from public vendor terms and documentation and are point-in-time. This document supports governance and compliance review; it is not a preventive security control.