On the surface, this week's Financial Times reporting — amplified by Simon Willison — is a market story: Anthropic's annualized revenue reportedly climbed to $65 billion in July (up from $47 billion in May), the company is telling investors it expects Q3 profitability, it claims roughly 6,000 customers spending $100,000 or more annually, and OpenAI's annualized revenue has jumped 35 percent to over $40 billion. Yet the headline tells the real story: Anthropic's best frontier model is struggling to attract users while cheaper tools thrive.
From a defender's chair, that sentence should set off alarms. When cost, not capability, drives AI adoption decisions, the security review process is usually the first casualty. Every dollar-driven switch to a cheaper, less-scrutinized AI tool is a potential new unmonitored channel for your organization's source code, customer PII, PHI, financial data, and credentials to leave your control.
After fifteen years of watching shadow IT evolve — from unsanctioned Dropbox accounts to rogue SaaS to today's AI explosion — I can tell you this pattern is familiar, but the velocity and data-sensitivity of shadow AI is unlike anything we've dealt with before.
Why "Cheaper Tools Thrive" Translates to "Ungoverned Data Flows Multiply"
The market dynamics described in this reporting create three concrete defensive problems:
1. Fragmentation of the AI vendor landscape. When premium frontier models lose ground to cheaper alternatives, employees don't wait for procurement — they self-select tools. Every new entrant in the AI market is another third party with its own data retention policy, training-data posture, subprocessors, and breach exposure. Your vendor risk inventory is almost certainly stale the week you finish it.
2. Race-to-the-bottom security postures. Cheaper tools achieve lower price points by cutting somewhere. Sometimes that's model size or inference cost — sometimes it's SOC 2 scope, data residency guarantees, encryption at rest for conversation history, or enterprise SSO/SCIM support. Consumer-grade tiers of AI tools frequently reserve the right to train on your inputs. If your developers are pasting proprietary code or your analysts are pasting incident timelines into a $10/month tool, you may be donating your intellectual property to a training corpus.
3. Invisible integration sprawl. The revenue numbers in this story — tens of billions in annualized spend — reflect how deeply AI APIs are now embedded in business workflows. Cheaper API pricing drives adoption of AI features inside SaaS products your teams already use. Features get enabled by default. Data starts flowing to model endpoints without a single ticket hitting your change management queue.
The Threat Scenarios We Actually See in IR
This isn't theoretical. In recent engagements, our team has investigated:
- Credential and secret leakage where developers pasted configuration files containing API keys and database connection strings into consumer AI chat interfaces to "debug faster."
- Regulated data exposure where healthcare staff used unapproved transcription/summarization AI tools on patient encounters — a HIPAA reportable event in waiting.
- Contractual violations where client data covered by NDAs and data processing agreements was processed by AI services with no DPA in place.
- Prompt-injection-driven exfiltration where AI agents with broad SaaS permissions (email, file storage, ticketing) were manipulated by malicious content in documents they were asked to summarize.
The common thread: none of these organizations had an inventory of which AI tools were in use, what data was flowing to them, or under what contractual terms.
Executive Takeaways
1. Build and maintain a live AI service inventory. You cannot govern what you cannot see. Use your CASB, secure web gateway, DNS logs, and expense reports to enumerate AI tools in use — both web interfaces and API integrations. Reconcile against your approved vendor list quarterly at minimum. Expect the list to be 3-5x larger than procurement believes it is.
2. Establish a fast-track AI vendor review process. The reason shadow AI wins is that official review processes take months while signing up for a tool takes ninety seconds. If your governance answer is "no, wait," the actual answer will be "yes, secretly." Create a tiered review: low-risk tools get 5-business-day triage, tools touching regulated or confidential data get full review with DPA, data-retention, and training-opt-out requirements.
3. Enforce data loss prevention at the AI egress points. Extend DLP policies to known AI service domains and API endpoints. Alert on — and where justified, block — uploads of source code repositories, documents tagged confidential, and bulk PII/PHI patterns to unsanctioned AI services. Pair technical controls with a clear acceptable-use policy so employees know the sanctioned path.
4. Contract for data protection, not just capability. For every approved AI vendor, verify in writing: no training on your data (or an explicit opt-out honored at the API level), defined retention and deletion timelines, breach notification obligations, subprocessor transparency, and data residency commitments. If a cheaper tool can't meet these terms, the price difference is your risk premium.
5. Govern AI agents and integrations, not just chat windows. The growing exposure surface in 2026 is AI features wired into email, file shares, CRMs, and ticketing systems via OAuth grants and API keys. Audit third-party OAuth grants in Google Workspace and Microsoft 365, scope AI service accounts to least privilege, and treat any AI agent with read access to internal data as a prompt-injection attack surface.
6. Educate the business on why "cheaper" can cost more. Market stories like this one — where budget tools win against premium models — are exactly what your CFO and department heads are reading. Get ahead of it. Explain that the security review exists to protect them from breach costs, regulatory penalties, and contract violations that dwarf any subscription savings. Give them an approved, cost-effective alternative and they'll rarely go rogue.
The Bottom Line
The financials in this story — $65 billion here, $40 billion there — confirm that AI adoption is not slowing down; it's commoditizing. Commoditization means proliferation, and proliferation without governance is how data leaves your organization through a thousand small, invisible doors. The defenders who win in this environment aren't the ones who block AI — they're the ones who make the sanctioned path faster, cheaper, and safer than the shadow path.
Security Arsenal helps organizations build practical AI governance programs, audit shadow AI exposure, and monitor for data egress to unsanctioned services. If your AI inventory hasn't been validated against actual network telemetry, that's the place to start.
Related Resources
Security Arsenal Penetration Testing Services AlertMonitor Platform Book a SOC Assessment vulnerability-management Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.