Viewpoints ·

Most AI data leaks run through approved tools. Approval isn’t control.

A new survey of 400 US and UK security leaders found that 47% have had sensitive data exposed through an AI tool, and in 77% of those cases the tool was one the company had approved. Fewer than half could say what kinds of data left. Approving a tool is a purchasing decision. Controlling what flows into it is a data job, and most firms haven’t funded it.

What happened

On 4 November, BlackFog published research it commissioned from Sapio Research: a survey of 400 IT security decision makers at companies with more than 500 employees, 200 in the US and 200 in the UK, fielded in September.

The headline findings:

  • 47% said their organization had an incident in which sensitive data was exposed through an AI tool.
  • 77% of those said a company-approved AI tool was involved.
  • 98% have an AI usage policy, but 35% rely on a written policy with no technical enforcement.
  • 62% could identify which users or departments send data to AI tools, but only 48% could produce any record of the types of data that left.
  • 56% said autonomous agents acting on employees’ behalf fall outside their existing controls, rising to 70% in technology and telecoms.
  • 39% said AI adoption is moving faster than they can govern it.

BlackFog sells data exfiltration prevention, so read the numbers as a vendor-commissioned survey. The pattern still matches the other cases this week.

My view

The finding that matters most is the 77%. The debate about AI risk has focused on shadow AI: staff using tools the company never approved. This survey says most exposure runs through the tools that went through procurement.

Approving a tool is not controlling the data

Approval answers one question: is this vendor acceptable? It does not answer the questions that decide exposure: what data may go into the tool, for which purposes, who may see the results, and where the prompts and outputs are kept. An approved assistant connected to a shared drive will read whatever the drive holds, including the folders nobody cleaned up, the spreadsheet with client data that should have been deleted and the document shared with “everyone” years ago. The tool is approved. The data it reaches was never reviewed.

You can’t protect what you haven’t labeled

Only 48% could say what kinds of data left through AI tools. That is a data hygiene finding, not a security tooling finding. If content isn’t classified, no control can tell a client file from a lunch menu, and the record of what left is a log of file names. The firms that can answer the question are the ones that labeled their sensitive data and made the label travel with it, so the control can act on the label rather than guess.

Agents widen the gap

More than half said agents acting for employees fall outside their controls. An agent inherits a person’s access and uses it continuously, across every system that person can reach. If that access was never tidied, because nobody removed old permissions or deleted what should have gone, the agent finds it all. Over-permissioned data was a slow risk when only people browsed it. With agents it becomes a fast one.

A written policy is a hope

A third of firms rely on a policy without technical enforcement. Policies matter, but a policy that nothing enforces describes what you hope people will do. The test I’d apply is simple: if someone pastes restricted data into an approved assistant today, what stops it, and what record shows it happened?

The same story, again

This is the third case this week with the same shape. At FNB, an access rule read a wrong field and nobody owned the field. At Microsoft, safeguards built for a smaller failure spread a larger one. Here, tools were approved without anyone owning what flows into them. Each time the missing piece is the unglamorous work: knowing what data you have, who owns it, what it may be used for and what can reach it. That work rarely has a budget line, and AI is making the cost of skipping it visible.

Why it matters

For the industry. Most enterprises buy AI inside products they already use, and those products reach data that was never classified or cleaned. Firms that treat AI exposure as a tool-approval problem will keep approving tools and keep leaking data. Firms that treat it as a data problem, label, own, minimize and enforce, will be able to say what left and why, which is what a regulator, a client or a board will ask after an incident.

How it connects to our work. Context Before Query sets out how an assistant gets approved meaning and permitted data before it answers. The Reference Implementation makes sensitivity labels travel with content and enforces them on every approved path, including to model providers. Patterns and Product Fit covers securing retrieval at enterprise and lean scale. Agent Onboarding gives each agent its own bounded identity instead of a copy of a person’s access. Six steps I’d take first:

  • Review what approved AI tools can reach. Not the vendor, the data: which drives, sites, mailboxes and databases each tool connects to.
  • Label sensitive data and make the label travel. Controls should act on the label, not on file names or guesswork.
  • Clean up before you connect. Remove stale sharing and over-broad permissions in any source an assistant or agent will read.
  • Enforce the policy in the path. Block restricted content from going to tools not approved for it, and log what was sent.
  • Keep a record of what left. By data type, tool and user, so you can answer the first question after an incident.
  • Give agents their own identity. Bound what each can reach, rather than letting it inherit everything its user can see.

This is one of three Viewpoints on why security, data hygiene and resilience are one problem. The others: A data error at FNB became a privacy breach and Microsoft’s outage started with a fix from July.

Sources

AISecurityData governanceAI agents

Linked sources are the record. Check them for the full reporting.

Read next

Viewpoint · · 6 min read

A data error at FNB became a privacy breach. Hygiene and security are one problem.

South Africa’s Information Regulator fined FNB R1 million after customers wrongly flagged as bank employees could open other people’s home loan applications. The access rule worked. The data it read was wrong, two warnings were closed as one-off tickets, and some records were kept with no stated reason. That is a data hygiene failure and a security failure at the same time, and AI agents will read the same flags.

SecurityData governance

News roundup · · 14 min read

The week in data and AI: California sues OpenAI over agents it lost control of

California sued OpenAI over at least 80 incidents in which its test agents escaped control, Cursor traced a trojanized Claude Code extension back to a key it had issued to Fireworks AI, and Amazon showed attackers paying for their AI agents with stolen AI keys.

AIData