Akamai: 6% of AI chats carry company data, 47% unmonitored

Six percent of your team's AI chats contain company data, and nearly half of those conversations are happening somewhere you can't see. Akamai published its State of the Internet security report, Speed, Scale, and Nonhuman Identity: The Agentic Threat Landscape, on September 22. Strip out the CISO framing and it is a set of numbers that describe the average small business more accurately than most small businesses would like.

What actually happened

Akamai's argument is that securing a modern business has shifted from identity and access management for people to behavioral governance over nonhuman entities — agents, extensions, and automated callers that hold credentials and act on their own schedule.

The findings that translate directly to a ten-person company:

• More than 40% of enterprise users have installed AI-powered browser extensions, and 25% of those extensions changed their permissions within 12 months of install.
• Those extensions are 60% more likely to carry known CVEs than standard extensions.
• More than 6% of enterprise AI chatbot conversations contain sensitive corporate data — mostly personally identifiable information — and 47% of those interactions happen through unmonitored personal accounts.

Akamai's recommendations run to edge-native runtime protections and API filtering, which is an Akamai-shaped answer to a real problem. But one of them is free and correct regardless of vendor: grant an agent autonomy in proportion to how verifiable its actions are and how reversible a failure would be, keeping a human in the loop for high-stakes decisions.

Why shadow AI matters for your business

A 25-person company has no CISO, no browser management policy, and a staff that installed a summarizer extension because it made Tuesday easier. Every number above describes you.

The permission-drift finding is the one people miss. An extension you vetted in March is not the extension running in September. It updated. It asked for more scope, got it silently through the browser's update flow, and now reads every page your bookkeeper opens — including the bank portal. Nobody did anything wrong. The review just had a shelf life nobody set.

What to actually do, in order of effort:

Run the extension inventory today. In Chrome or Edge, have each person open the extensions page and screenshot it. You are looking for anything that reads data on all sites. Twenty minutes, no tooling, and it usually finds two or three things nobody remembers installing.

Give the team a sanctioned AI account so they stop using personal ones. The 47%-on-personal-accounts number is not a discipline problem; it is a provisioning problem. People use the tool that's in front of them. A paid workspace account with admin visibility costs less than one incident and removes the incentive entirely.

Write down what's not allowed in a chat window. Client PII, credentials, contract terms, anything under an NDA. One page. The rule only works if it exists before someone needs it.

Scope agent credentials to reversible actions. Akamai's autonomy-proportional-to-reversibility rule is the right default. Read-only by default. Draft, don't send. Stage, don't publish. If an agent can do something you cannot undo in one click, it needs an approval step — not because the model is bad, but because you have no way to verify it was right.

Sources: Akamai press release, Akamai State of the Internet.