Companies are pouring time and dollars into security awareness training, but there is little empirical evidence to suggest it actually works against social engineering.
Social engineering remains a primary and successful attack vector. While system vulnerabilities can be patched, social engineering cannot. The most common pseudo ‘patch’ is user awareness training, but this has failed to block the vector. Human defenders should not and cannot be expected to detect trickery designed to manipulate their psychology. And the tricks are becoming better hidden and more sophisticated with the use of AI deep fakery.
Successful examples of social engineering include:
The Las Vegas casino breaches of 2023, via vishing. Scattered Spider attackers, posing as employees, tricked the casino’s help desk staff to reset passwords and bypass MFA. MGM Resorts was forced to shut down digital operations for ten days at an estimated cost of $100 million in lost revenue and remediation costs. Caesars Entertainment is thought to have paid a $15 million ransom after the attackers demanded $30 million.
More recently, the Brinks Home leak disclosed in August 2026 also involved voice phishing and the enterprise help desk. This time the attacker was ShinyHunters, although the group has a known relationship with Scattered Spider. Here the attacker simply talked a help desk employee through completing a Microsoft Entra authentication step, granting the hacker immediate access. Almost five million customer records and 41 gigabytes of corporate data were subsequently published on a public hacking forum.
If human detection of social engineering cannot be assured, other means of detection must be employed – and technology is the pre-eminent option. The use of technology to detect both traditional and newer AI deep faked social engineering is increasing and improving. One example comes from Netarx.
The firm’s solution applies a proprietary orchestration layer which it describes as an AI defense meta harness. The harness correlates individual signals detected by more than 40 AI models continuously scanning all communications while they are in progress, analyzing more than 1,000 digital and metadata signals for each interaction.