ChatGPT web-page flaw highlights new phishing risk for local businesses
You and me go ChatGPhish-ing in the dark
By The Register
A newly reported ChatGPT security issue has highlighted why businesses using AI tools should be cautious when asking them to summarise web pages, emails or online documents.
The Register reported that a researcher found a prompt-injection technique in which malicious instructions hidden inside web content could be treated by ChatGPT as part of the task. In simple terms, a web page could contain instructions that are not obvious to the user but may influence how the AI responds.
The issue has been described by security researchers as ChatGPhish because of the way it could potentially turn web-page summaries into a phishing risk. The concern is not that every ChatGPT user is automatically affected, but that AI tools can sometimes struggle to separate trusted user instructions from untrusted material pulled in from websites.
For local businesses in Cheshire, the practical lesson is straightforward: AI tools are useful, but they should not be treated as a fully trusted security layer.
Many small firms now use AI to summarise supplier pages, compare services, draft emails, review technical documents or make sense of long online guidance. That can save time, but it also creates a new risk if the source material itself contains misleading instructions, fake links or hidden prompts designed to influence the AI response.
A member of staff may ask an AI assistant to summarise a page and then trust the answer without checking the original source. If the page has been manipulated, the AI could potentially present unsafe links, misleading contact details or a convincing but false instruction.
This is particularly important for businesses handling invoices, customer data, login details, supplier accounts or payment information. Phishing attacks already rely on urgency and trust. AI can add another layer of confidence if a user believes the answer has been independently checked by a smart tool.
The safest approach is to use AI as a helper, not as final proof. Businesses should still check links manually, verify bank details through known contacts, avoid pasting sensitive information into public AI tools, and be cautious with any AI-generated instruction involving payments, passwords or account access.
Staff should also be trained to understand that prompt injection is different from traditional malware. A page does not necessarily need to install anything on a device to cause problems. The risk can come from text or formatting that attempts to steer the AI tool’s answer.
For business owners, this means AI policies should now sit alongside normal cybersecurity rules. A simple internal rule could be: never rely on an AI-generated link, phone number, payment instruction or login route unless it has been checked against a trusted source.
The report is another reminder that AI adoption is moving faster than many organisations’ security habits. Tools such as ChatGPT can improve productivity, but they also require judgement, especially when they interact with web pages and external content.
Cheshire businesses using AI for admin, customer service, marketing or research should review how staff are using these tools and make sure basic verification steps remain in place.