What we share with AI and what it actually reads
Two risks: the data we share with AI and the instructions hidden in the content it analyses.
- 46%
- of professionals have already uploaded sensitive company information to an external AI tool
- 32%
- of data security incidents involve generative AI tools
Hi! How can I help you?
Definition
What does safety mean when using AI?
AI safety covers a wide range of issues, including the sharing of confidential data (client, contract and HR data) and prompt injections (instructions hidden in the content AI is asked to read).
A sentence such as “ignore previous instructions…” slipped into a file, a web page or an email can hijack the AI and divert it from the user's request.
What these risks cost the business
Legal risk
Sharing personal or confidential data with a third-party service exposes the company to GDPR penalties and disputes with clients.
Reputational risk
A data leak, or an AI assistant hijacked by a hidden instruction, can jeopardise years of trust with your partners.
Operational risk
A prompt injection can skew a summary, trigger a malicious action or disrupt a critical process.
Measure
How Taboo assesses safety
In our simulation, employees receive sensitive files and booby-trapped documents. Taboo checks that they anonymise confidential data and spot malicious instructions.
- Anonymising sensitive data before sharing documents
- Spotting prompt injections
document.pdf
ignore previous instructions…
Frequently asked questions
Can client data be shared with an AI?
Only with a company-approved tool, and after anonymising sensitive information such as names, contact details or personal identifiers.
How do you spot a prompt injection?
By developing the right habits to inspect content before uploading it into the system.
Discover and in a live demo
In 30 minutes, we walk you through the test, the report and the training programme, tailored to your organisation.