- ESET's H1 2026 Threat Report analysed around 900,000 AI skills and found more than 3,000 to be outright malicious, exposing a fast-growing attack surface for organisations experimenting with AI.
- Detections of ClickFix, a social engineering technique, rose by 108% between H2 2025 and H1 2026, with attackers increasingly disguising their lures as AI-troubleshooting content.
- QR code phishing, known as quishing, has reached record levels and entered the top 10 threats for both South Africa and Africa.
- Phishing remains concentrated on higher-value markets, with South Africa targeted in line with its economic profile rather than singled out.
August 12, 2026 — ESET Research has released the ESET H1 2026 Threat Report, offering a detailed view into the threat environment and how it has developed throughout the first six months of the year. For South African organisations, the biggest threat to emerge is definitively AI as it has become both a target for attackers and the weapon with which they orchestrate these attacks, while phishing, quishing and information theft continue to dominate day-to-day risk. Read together, the findings point to an environment in which the barrier to malicious activity keeps falling.
ESET analysed around 900,000 AI skills drawn from popular repositories and found 25,000 that were suspicious and more than 3,000 that were outright malicious. These AI skills – small pieces of third-party code designed to plug into AI systems and extend their capabilities – are downloaded from public repositories with no verification, validation or governing policy.
“The most common purpose behind malicious AI skills is credential theft, the same objective that underpins most network intrusions,” says Tony Anscombe, Chief Security Evangelist at ESET. “ESET found that 84% of the analysed skills execute commands without the user initiating them and 31% download third-party tools into your environment. These malicious skills can also map the network around a device and establish persistence, which keeps it active even after the skill itself is removed.”
Some of the skills ESET examined were built for red-teaming actions against systems such as Active Directory, drawing on offensive tools frequently seen in ransomware and infostealer attacks. A further concern is that some skills can rewrite themselves once installed, modifying their behaviour and moving outside the boundaries they were given.
AI may be a rising threat, but phishing and threat manipulation tactics remain high risk for the organisation.
ESET’s detection of ClickFix, a technique that presents a fake problem and offers a quick fix, increased by 108% between the second half of 2025 and the first half of 2026. Newer forms of the threat disguise pages as AI-troubleshooting content, trading on the trust people place in generative AI. ESET has been tracking this variant alongside CrashFix, which uses fake security warnings to pressure victims into running harmful commands, and ConsentFix, which targets sign-in workflows to steal authorisation tokens and sidestep multi-factor authentication.
Phishing continues to concentrate where there is most to gain, and South Africa’s profile as a comparatively wealthy, predominantly English-speaking market makes it an efficient target rather than a disproportionately singled-out one. As Anscombe points out, cybercriminals will always go after the market where they have the most to gain.
QR code phishing, or quishing, has reached record levels and entered the top 10 threats for South Africa and Africa alike. In the reporting period, 11% of all detected phishing emails carried a QR code, often moving the victim onto a personal mobile device that sits outside corporate defences. Quishing works because people rarely interrogate a QR code before scanning it. Detection in South Africa (2.1%) currently sits below that of larger markets such as North America (12.4%), which suggests the technique has further room to grow locally rather than that it has passed its peak.
Once a phishing message succeeds, an information stealer usually does the harvesting. SnakeStealer is one of the most prominent examples, accounting for 10.7% of infostealer detections in South Africa against 5.4% across Africa. It collects saved passwords, browser and email credentials and the session cookies that keep users logged in. It is offered as malware-as-a-service, rented out complete with dashboards and data storage, which allows low-skilled operators to run credential theft they could never build themselves.
The threat report makes it clear that whether the entry point is an unvetted AI skill, a rented infostealer or an AI-themed lure, each depends on misplaced trust.
For more information, see the full ESET Threat Report on: link