One in five UK workers has handed a password, API key or login credential to an AI tool or agent, according to research published this week by the password manager Dashlane. Nearly half have let an AI system act on their account without asking anyone first, including reading email, opening company files and pulling customer information.
Misplaced Trust: Inside the Workplace's New AI Security Blind Spot, published this week, surveyed 1,000 UK workers and 250 UK IT executives as part of a wider group of 3,800 across the UK, US and France. Its central finding is not that people are careless in the abstract. It is that they extend trust to software the way they extend it to a new colleague, before that trust has been earned.
What the survey found
- 19 per cent have shared a password, API key or login credential with an AI tool or agent
- 48 per cent have allowed AI to perform at least one activity without seeking permission, including accessing emails, company files and customer information
- 37 per cent would lend their login credentials to a new AI tool after one week of using it
- 23 per cent have pasted sensitive information into an AI tool that they would not share with a colleague, a website or another work system
- 21 per cent did not check where that information would be stored or sent
- 15 per cent granted an AI tool access to a file, system or account it turned out it should not have reached, and 45 per cent of those say the access is still active
Only a quarter of workers say they always consider security risk when using an AI agent at work.
Why it matters more than ordinary password reuse
A shared password sits in a database that can be rotated. A credential handed to an agent can act, and can keep acting after the employee has forgotten the conversation.
Dashlane calls the pattern the "AI Colleague Effect". Its chief product officer, Christophe Frenet, said employees are "overextending trust to AI systems, sharing credentials and granting access to AI agents, sometimes without fully understanding what those tools can do with the permissions they've been given". That, he said, creates real difficulty for security teams trying to detect misuse or investigate an incident.
The gap between what staff do and what IT believes is the sharpest part. Seventy-two per cent of the IT executives said their organisation had suffered an incident involving an AI agent's credentials or access in the past 12 months. Yet 92 per cent were confident they know about every AI agent operating inside their organisation.
An agent given mailbox access can read a password reset email. One given file access can copy a document to a third party. Neither needs to be malicious, and neither looks like a person logging in at an odd hour.
What to do this week
The remedy is unglamorous and mostly already exists.
List the agents with access. Every AI tool connected to a work account has a permissions page. Google, Microsoft and most SaaS products show which third-party apps can read mail, files or calendars. Revoke anything nobody recognises.
Stop pasting credentials into prompts. An API key typed into a chat window is a key in someone else's logs. Give a tool a scoped token that can be revoked, not the master credential.
Apply the rules you already use for people. Frenet's summary is the right one: just-in-time provisioning, least privilege, credential hygiene and proper authentication have to apply to AI colleagues too. Passkeys and scoped tokens do most of that work.
Treat the audit as continuous. The 45 per cent whose unwanted access stayed active were not attacked. They never went back to check.
For anyone who has already pasted a password somewhere it should not be, start with how to check whether your password has leaked, then move the account to a manager from our pick of the best password managers for 2026. If you want to remove the password entirely, what a passkey changes explains the switch. The same instinct about what software retains applies at home, as Ofcom's file on smart speaker data retention sets out.
The ICO's guidance on AI and UK GDPR is the reference for what an employer is allowed to do with data an agent touches.
The bottom line
The finding is not that UK workers are reckless. It is that AI tools arrived with the permissions of a trusted colleague and the oversight of a browser tab. One in five people have handed over a credential, and most will not remember doing it. The fix is a permissions review, not a policy document.

