In brief: Cybercriminals are increasingly attacking AI-based protection mechanisms directly, while AI-powered website builders and OAuth-based consent phishing open new attack surfaces.
Proofpoint documents that attackers are directing their methods specifically against the AI-based detection systems that companies rely on. At the same time, AI tools are dramatically lowering the barriers to entry for professional phishing campaigns.
Proofpoint’s latest report shows a paradigm shift in cyber defence: rather than exploiting technical vulnerabilities, attackers are targeting the AI systems that companies deploy for protection. This turnaround undermines the previous assumption that AI-based security solutions represent an effective countermeasure to AI-enabled attacks.
In so-called prompt injection attacks, criminals hide invisible instructions in phishing messages that are intended not for the recipient but for the analysing AI instance. These instructions prompt the systems to, for example, undergo deep multi-layer reasoning loops, generate at least ten different internal perspectives, or recursively refine thoughts. The goal: to overload detection systems with compute-intensive requests so that malicious messages remain undetected due to timeouts.
Simultaneously, the creation of professional phishing campaigns is becoming significantly easier. Proofpoint documented a case in which criminals used the AI website builder Lovable to create a deceptively authentic YouTube complaint portal – without requiring any programming knowledge. Victims were directed to a supposed copyright complaint process in which malware was installed using so-called ClickFix techniques, which often bypass traditional protection mechanisms. According to Proofpoint, hundreds of thousands of malicious websites are created monthly via such AI platforms.
A third critical point concerns the exponential increase in AI applications with OAuth authorisation. The number of such applications rose within a year from approximately 11,000 to over 258,000. This development drives so-called consent phishing attacks, in which users are tricked into granting seemingly trustworthy applications broad access rights – often without critically examining the requested permissions. In German companies, the risk is particularly relevant: 77 percent are already testing or using autonomous AI agents based on such permission structures. Violations can have regulatory consequences beyond technical implications.
Source: www.it-daily.net · Published 8 June 2026
Lumi AI News — AI-assisted curation according to Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.6.5.