November 20, 2026

AI can now fake your face better than experts can spot it, new report warns

Check-Point-AI-report

If you think you could spot an AI-generated face, the odds may not be in your favour. A new report from Check Point Research found that even people trained specifically to detect deepfakes correctly identified AI-generated faces only about 41% of the time. Ordinary viewers did worse, catching just 30%.

The finding is one of several in Check Point’s Annual AI Security Report 2026, an assessment of how artificial intelligence has changed the cyber threat landscape over the past year. Its overall conclusion is stark: AI has stopped being just a tool that helps cybercriminals work faster and has started actively carrying out attacks with very little human involvement.

When a face, a voice or a video call can’t be trusted

The report argues that the traditional ways we prove who we are online, a familiar voice on the phone, a face on a video call, a scanned ID, are no longer reliable on their own. Voice-cloning tools can now convincingly copy someone’s voice from a short audio clip, and are already being sold as ready-made fraud services. One platform identified by researchers uses AI voice agents to talk victims through fake account-recovery calls at scale, with no human caller involved.

Video isn’t much safer. The report describes real-time face-swapping now being used on live video calls by both state-linked hacking groups and organised scam operations, including calls in which criminals used deepfake video and hijacked messaging accounts to trick people into handing over cryptocurrency. In one investment scam documented by Check Point Research, victims were lured into WhatsApp and Telegram groups populated almost entirely by AI-generated personas posing as fellow investors, complete with fake profile photos and fabricated trading results, none of it real.

Separately, the report links a large-scale North Korean operation to the use of AI-generated identities, resumes, and interview personas to get operatives hired as remote IT workers within Western companies, generating close to $800 million (£630m) for the regime, according to US Treasury sanctions cited in the report.

AI is now doing the hacking, not just helping with it

Perhaps the most striking case in the report is a breach affecting nine Mexican government agencies between December 2025 and February 2026, which exposed roughly 400 million records, including tax, medical, and electoral data. Researchers found that a single attacker typed just 1,088 instructions, while AI tools carried out 5,317 separate attack commands on their behalf across more than 30 sessions.

When the AI assistant initially refused to help with anything malicious, the attacker simply pasted hacking instructions into a configuration file the tool automatically trusts every time it starts up, permanently bypassing its safety rules without needing to ask again.

The report says this pattern, criminals using commercial AI tools such as ChatGPT, Gemini and Claude, deliberately steering around their safety features, is now the most common way attackers access AI capability, more common than building their own AI tools or buying illicit versions on the dark web.

Businesses are leaking their own data, often by accident

The report also flags a quieter, more everyday risk: employees sharing sensitive company information with AI chatbots without realising the danger. Check Point’s data shows that the proportion of “high-risk” prompts, meaning ones containing sensitive company, personal or regulated information, doubled over the past year, from 2% to 4% of everything typed into AI tools. Between 87% and 93% of organisations had at least one high-risk AI interaction every single month.

The average business now uses about ten different AI applications a month, the report finds, often without any formal sign-off from IT or security teams, a trend researchers describe as adoption outpacing governance.

What Check Point is saying needs to change

Lotem Finkelstein, Vice President of Check Point Research, said the shift marks a turning point from previous years, when AI was mainly viewed as a way to make existing hacking techniques faster and cheaper.

“AI has crossed into the live attack chain,” Finkelstein said, warning that defenders can no longer assume a human is setting the pace on the other side.

The report calls for organisations to rethink identity verification, monitor how staff are actually using AI tools day to day, and treat the security of their own AI systems, not just the data those systems handle, as a standing priority rather than a box to tick once during rollout.

Check Point’s Annual AI Security Report 2026 is available now here: https://engage.checkpoint.com/ai-security-report-2026