The awkward wording and amateur formatting that once gave away a scam are disappearing, as fraudsters use AI to produce convincing imitations of SARS correspondence ahead of the 23 October tax filing deadline.
Kaspersky researchers found scammers had built a fake SARS eFiling page that looks and functions like the official version, asking victims for usernames, passwords and phone numbers. A separate scam presents a counterfeit Adobe PDF sign-in screen requesting email credentials.
“Some of the old visual clues, including awkward wording or amateur-looking correspondence, no longer have the same impact,” Kaspersky said. “A message can look professional and still be fraudulent.”
SARS itself warned in July that scammers were using AI to generate “professional-looking email templates that are harder to identify as fraudulent.”
Why the timing works for fraudsters
The deadline is part of the design. Pressure makes people easier to manipulate, and by late in the season, communications from SARS are routine.
“By this point in the tax season, people are used to receiving messages about returns, assessments and refunds, so an email like this may not immediately look out of place,” said Chris Norton, Kaspersky’s general manager for sub-Saharan Africa.
The messages typically promise a large refund, or threaten a court summons over unpaid tax or alleged evasion. Links lead to fraudulent sites that harvest credentials.
The other risk: asking a chatbot for help
Kaspersky’s second warning concerns legitimate AI tools rather than malicious ones.
Taxpayers increasingly turn to chatbots to understand their returns or work through tax questions. Kaspersky advises against uploading tax records, identity documents, banking details or other confidential information to any public AI service — both because the responses may contain errors and because the data may not stay private.
The concern is grounded. In 2025, more than 370,000 conversations from xAI’s Grok were indexed by Google and Bing because the platform’s share feature generated publicly crawlable links. The exposed chats included private medical and psychological information and business strategy documents. Most major platforms — including Anthropic’s Claude, OpenAI’s ChatGPT and DeepSeek — have experienced comparable exposure through sharing features.
“AI can make complicated information easier to work through, but taxpayers should be careful about the information they provide to an external service,” Norton said. “Sensitive records should stay out of public AI tools, and important tax information should be checked through trusted sources.”
What to do
- Go to SARS directly. Type sars.gov.za into your browser rather than following any link in an email or SMS.
- Treat refund offers and summons threats as suspect by default, however professional they look.
- Never enter eFiling credentials on a page reached through a link.
- Keep tax documents out of public chatbots. If you use AI to understand a tax concept, describe the question in general terms without uploading records or personal details.
- Verify with SARS through its official contact channels if unsure.
A wider pattern
The tax scams are one instance of a continental trend. INTERPOL’s African Cyberthreat Assessment Report 2026 attributed 55% of reported cybercrime on the continent to AI-enabled methods, with losses more than doubling to $484 million. South Africans have already been targeted by a deepfake of Reserve Bank Governor Lesetja Kganyago directing victims to fraudulent investment platforms.
Boston Consulting Group estimated this month that agentic AI could cut the cost of running scams by 90% or more — lowering the barrier enough to draw in new operators who previously lacked the skill to run convincing campaigns.
The practical consequence for taxpayers is simple. Whether a message looks right no longer tells you whether it is.
