Akamai today emailed marketers and publishers a promotion for its Enterprise AI Usage Risk Report 2026, a 27-page special edition of its State of the Internet series dated August 2026. The report finds that 47.11% of enterprise AI conversations run through personal identities rather than corporate-managed accounts. The numbers are not new: they come from LayerX, a browser security company Akamai acquired, and match the State of AI Usage Report 2026 that PPC Land covered in May.

In Short

Akamai, a company that sells internet security products, put out a report saying that almost half of the conversations workers have with AI chatbots happen on personal accounts their employer cannot see. That matters to any company whose staff paste customer lists, budgets or campaign plans into ChatGPT or similar tools, because data in a personal account sits outside corporate controls. The figures are recycled from research by LayerX, a company Akamai bought, and the new version leaves out the data point that actually measured how often sensitive data is shared.

What Akamai sent and what it published

The email arrived at 15:02 on October 21, 2026, from Akamai Technologies under the subject line "Report: Employee AI usage raises risk." It opens with a question about staff using "a personal favorite AI tool for work" and then states: "Industry research shows 47% of enterprise AI conversations run on unmanaged personal accounts." The email does not name the research behind the number. It promises four highlights: the top five enterprise AI usage risks, an explanation of "how copy/paste activity and high-risk prompts leak sensitive data," the percentage of employees running AI extensions with critical permissions, and a checklist for chief information security officers.

The report itself is labelled State of the Internet, volume 12, issue 04, 2026. Its back page says "Published 08/26," and the PDF file was created on July 28, 2026. So the email is today's news, while the document it promotes is about two months old. Kimberly Gomez is credited as research director, Maria Vlasak with writing and editing, and Eyal Arazi and Pooja Gupta with review and subject matter contribution.

Akamai sells security products. The report's closing boilerplate describes Akamai Security as protecting "the applications that drive your business at every point of interaction," and most of the final third of the document is a mitigation framework and checklist that maps closely onto browser-level AI monitoring. The findings are, in that sense, vendor claims made by a company with products to sell in the category it describes.

Where the numbers come from

The provenance is stated once, in the introduction: "With Akamai's recent acquisition of LayerX, we received access to their data." Every quantitative finding in the report is then attributed to LayerX, and the hyperlinks embedded in the PDF point to LayerX's State of AI Usage Report 2026 on the layerxsecurity.com library. The report does not give a date for the acquisition.

That LayerX report is not new to PPC Land readers. On May 31, 2026, PPC Land reported that nearly half of enterprise AI runs on personal accounts, based on the same LayerX telemetry. Set side by side, the figures match to the second decimal place: 47.11% personal, 52.89% corporate, 18.24% weekly users, 5.09 prompts per conversation, 16.31% of AI extensions with known CVEs. Akamai's version is a repackaging of that dataset under its own State of the Internet brand, with added case studies and a mitigation section.

What neither document provides is a method. Akamai's report gives no number of organisations, no number of users or conversations, no collection period and no geography. The only definitions are the company-size bands used for the extension figures: small enterprises under 1,000 employees, medium-sized enterprises between 1,000 and 2,500, and large enterprises above 2,500. PPC Land's May coverage found the same gap in LayerX's original, which described its source only as telemetry from its browser security platform inside organisations of varying sizes. A separate LayerX extension study, covered by PPC Land in May, did cite enterprise telemetry from more than 1 million devices, but Akamai's report does not say whether the extension figures it reproduces come from that base.

The identity findings

The headline finding is the split between personal and corporate identities. According to Akamai, 47.11% of enterprise AI conversations "take place through personal identities rather than corporate-managed accounts," against 52.89% through corporate accounts. Akamai frames this as shadow AI: "business activity occurring in AI environments that IT never approved, security teams can't monitor, and compliance teams can't audit."

A second figure concerns a narrower group. Among conversations that used corporate identities, 14.4% were tied to personal licences rather than corporate or business licences, according to the pie chart labelled Figure 6. The body text describes the same number differently, saying 14.4% "of enterprise AI conversations occurred using corporate email addresses linked to personal 'freemium' AI subscriptions." The chart's base is corporate-identity conversations, not all enterprise conversations, and those two bases produce different totals. Read by the chart, the share of all conversations would be about 7.6%, a PPC Land calculation from 14.4% of 52.89%.

The report then makes a claim it does not support with any data. Of those corporate-email accounts on personal subscriptions, it states that "the data going into those accounts was being used by the LLM for public model training." The report does not say which providers were involved, whether training settings were checked, or how LayerX could observe what a provider does with data after it is submitted. Whether a consumer LLM account feeds model training depends on the provider and on the user's settings, and the report treats a possibility as an observed fact.

Platform by platform

Figure 7 breaks the identity split down by product. The chart's values are:

  • ChatGPT: 38.64% corporate, 61.36% noncorporate
  • Gemini: 62.15% corporate, 37.85% noncorporate
  • Gemini Enterprise: 98.15% corporate, 1.85% noncorporate
  • Claude: 38.91% corporate, 61.09% noncorporate
  • Copilot: 36.08% corporate, 63.92% noncorporate
  • Copilot M365: 90.55% corporate, 9.45% noncorporate
  • DeepSeek: 0.17% corporate, 99.83% noncorporate

The accompanying text says ChatGPT, Copilot and DeepSeek "are primarily accessed through personal accounts, with more than 60% of conversations tied to personal identities." Claude, at 61.09% noncorporate in Akamai's own chart, also clears that threshold and is not named. PPC Land's May article on the LayerX original listed Claude alongside the other three. The omission does not change any number, but the text and chart disagree on which products belong in the group.

DeepSeek gets a dedicated case study. Akamai describes how US policymakers in 2026 intensified scrutiny of the Chinese company, "with proposals ranging from government restrictions to adding the company to the U.S. Entity List," linking to a Tom's Hardware report and to a CISA page. The chart shows almost no DeepSeek use through corporate accounts at all.

Usage concentrates among a small group

The report's first risk is the expansion of the threat surface through what it calls AI power users. According to Akamai, the average enterprise user takes part in 36 AI conversations, the bottom 50% engage in 12 or fewer, and the top 5% generate at least 144. The average conversation contains 5.09 prompts, while the top 5% of conversations contain at least 18.

There is a labelling error in Figure 2. The data label on the 50th percentile point reads 2, while the caption and body text both say 12. The plotted point sits near 12 on the axis, so the label appears to be wrong rather than the text. Figure 3, the prompts chart, also labels its median as 2, which is consistent with the one to two prompts per median conversation reported in May.

The frequency figures need care. Figure 1 shows 18.24% of employees writing AI prompts weekly, 30.47% monthly, 40.57% quarterly and 47.67% in a "Quarterly+" column. These add up to well over 100%, and they rise as the window lengthens, which suggests each figure counts users active at least once in that period. The report does not say so. Its text adds to the confusion by stating that "almost 50% use AI more casually," which could refer to the 47.67% column or to a residual group; the report does not specify.

Applications follow a similar curve. The report says the top four AI applications in an organisation are each used, on average, by more than 20% of staff, falling below 5% after the tenth application and close to zero by the thirtieth. Figure 4 carries no data labels, and its curve drops below 5% at around the eighth application, so the chart and text describe the tail slightly differently.

The section that has no numbers

The email promises a look at "how copy/paste activity and high-risk prompts leak sensitive data." The report's data leak section contains no figures at all. It lists the channels through which data now moves, including prompts, conversational context, source code snippets, screenshots, copied text and generated responses, and argues that "the biggest AI security risk is no longer employees who access AI, it is employees who share sensitive business data with AI." Copy/paste appears only in the mitigation framework and checklist, as an activity to inspect.

The LayerX original did measure this. As PPC Land reported in May, LayerX found that 6.48% of all enterprise AI conversations contained sensitive data, with personal data in 5.81%, financial data in 0.96% and IT or security data in 0.94%, and it gave sensitive-data exposure rates per platform, from 3.65% for Copilot M365 to 12.63% for DeepSeek. None of those figures appear in Akamai's edition. The report that Akamai markets on data leakage leaves out the one measurement in its source dataset that addressed leakage directly.

In their place, Akamai offers a case study. It describes how, in early 2026, "a U.S. government official accidentally released internal, restricted operational data" from the Cybersecurity and Infrastructure Security Agency through a public AI service. The report does not name the official; its embedded link points to a Politico article dated January 27, 2026, whose URL refers to CISA's Madhu Gottumukkala and ChatGPT. Akamai says the exposed material included government documents meant for internal use and data that was not classified but was not intended for public AI systems, and stresses that the incident "was not the result of a cyberattack or technical compromise."

Extensions: the email's promise and the report's data

The extension findings are the most detailed in the report, and they are also where the email overreaches. According to Akamai, 14.55% of users at small enterprises, 17.70% at medium-sized enterprises and 9.53% at large enterprises run at least one AI browser extension. The report attributes the lower large-company figure to organisations that "remain more conservative or regulated," an interpretation offered without supporting data.

Permissions are measured across installed AI extensions, not across users. According to Akamai, 56.4% of AI extensions request high permissions, 16.7% critical, 25.6% medium and 1.4% low. The report rounds high and critical together to "nearly 75%"; the sum is 73.1%. The email promised "the percentage of employees running AI extensions with critical permissions." The report contains no such figure. It gives the share of employees running any AI extension and the share of extensions requesting critical permissions, and the two cannot be combined without knowing how many extensions each user runs.

The comparison with ordinary extensions is the clearest finding:

  • 16.31% of AI extensions have known CVEs, against 10.8% of all extensions
  • 18.19% of AI extensions request cookie access, against 6.67% of all extensions, which the report describes as "nearly three times" the rate
  • 41.91% of AI extensions request scripting access, against 15.4% of all extensions

These match the figures LayerX published in its April extension report, which PPC Land described as showing AI extensions 60% more likely to carry a known CVE.

The report warns that an extension with elevated permissions "could easily expose sensitive information or take over user sessions." Ad platforms, analytics suites and CRM tools are typically used through logged-in browser sessions, which is where those permissions apply. The report does not break the data down by department or by extension category, so it gives no direct view of marketing teams specifically.

CursorJacking and vibe hacking

Two LayerX research projects illustrate the extension and coding-tool risks. In CursorJacking, LayerX showed that a rogue browser extension could extract Cursor API keys, which Cursor stores in a local SQLite database. Akamai's diagram shows sample keys for OpenAI and Anthropic and a session token being pulled from that database and sent to an attacker-controlled server.

The vibe hacking section describes LayerX's 2026 demonstration that "a popular frontier AI coding assistant" could be manipulated by modifying a project's local markdown instruction file. The report does not name the product. Its embedded hyperlink does: the link points to a LayerX blog post whose URL says Claude Code "can be turned into a nation-state-level attack tool with no coding at all." That detail is visible only in the link annotation, since text extraction drops it.

Agents and AI browsers

The fifth risk concerns agentic AI systems that can take actions inside corporate software. Akamai's example is CometJacking, which it dates to 2025: LayerX researchers showed that a crafted link could instruct Perplexity's Comet browser to collect data, encode it in Base64 to get past security checks, and send it to a remote server. Akamai describes this as an indirect prompt injection attack. PPC Land covered CometJacking alongside Brave's findings on Comet in October 2025, noting that LayerX identified the issue on August 27 and 28, 2025, and that Perplexity responded that there was "no security impact."

The report offers no telemetry on agent use. The LayerX original had one relevant figure, which PPC Land reported in May: only 0.16% of enterprise users ran at least one connector-enabled AI workflow. Akamai's edition does not include it, though it describes agents as "fast becoming the primary gateways through which users interact with corporate systems."

Recommendations in the report

Akamai closes with a mitigation framework aimed at CISOs, organised around five priorities: focusing on power users, eliminating shadow AI, protecting data at the interaction layer, securing extensions and development environments, and preparing for autonomous agents. The checklist includes enforcing single sign-on and multi-factor authentication for AI applications, monitoring "personal-to-corporate account crossover," and inspecting prompts, uploads and copy/paste activity in real time. Most of these measures require software that sits in the browser and reads what employees type into AI tools, which is the category LayerX sold before the acquisition.

Why this matters for marketing teams

Marketing departments are among the heaviest users of the tools in Akamai's charts. Basis's annual survey, published this week, found ChatGPT used by 74.5% of 184 marketing respondents and Claude by 72.3%. In April, ProGEO.ai's survey of 112 marketing professionals found 76% using generative AI daily while only 43.8% said technical controls enforced their company's policy, and 57.1% paid for a platform personally. Those are the conditions under which the personal-account share in LayerX's data becomes a practical question about client briefs, customer lists and media budgets.

The consumer tools themselves have also been found to leak. Researchers at IMDEA Networks reported this month that 6 of 9 AI chatbot web clients passed conversation URLs, titles, prompts or screenshots to third parties. In September, a CCC and Outsell survey of 570 knowledge workers found employees putting externally published content into AI tools about 11 times per week, with executives carrying roughly six times the copyright exposure of individual contributors. And browser extensions are gaining new powers: OpenAI's ChatGPT Chrome extension can now operate open tabs to fill forms and send messages, the kind of capability that the report's permission figures concern.

Against that backdrop, Akamai's report adds case studies and a governance checklist, but no new measurement. Its strongest numbers were public in May. Its weakest claims, on model training and on copy/paste leakage, are the ones its marketing email leads with.

Timeline

Summary

Who: Akamai Technologies, the security and content delivery company, using data from LayerX, a browser security company it acquired. The report credits Kimberly Gomez as research director.

What: The Enterprise AI Usage Risk Report 2026, a 27-page State of the Internet edition. It finds 47.11% of enterprise AI conversations run through personal identities, 14.4% of corporate-identity conversations use personal licences, and AI browser extensions carry known CVEs at 16.31% against 10.8% for all extensions. The figures match LayerX's State of AI Usage Report 2026 from May. The report states no sample size or collection period, leaves out LayerX's sensitive-data measurements, and contains a chart labelling error and an unsupported claim about model training.

When: The report is dated August 2026. Akamai promoted it by email today, October 21, 2026.

Where: Enterprise environments monitored by LayerX's browser security platform. The report gives no geography or organisation count.

Why: Akamai argues that personal accounts, unreviewed apps, extensions and agents put corporate data outside security controls, and it recommends browser-level monitoring of AI use, a category in which it now sells products. For marketing teams that rely heavily on consumer AI tools and browser extensions, the identity and extension figures describe a real exposure, though the numbers are vendor claims and not new.