Google urges caution as Gmail AI tools face new threats

Google has issued a warning about a new wave of cyber threats targeting Gmail users, driven by vulnerabilities in AI-powered features.

Researchers at 0din, Mozilla’s zero-day investigation group, demonstrated how attackers can exploit Google Gemini’s summarisation tools using prompt injection attacks.

In one case, a malicious email included hidden prompts using white-on-white font, which the user cannot see but Gemini processes. When the user clicks ‘summarise this email,’ Gemini follows the attacker’s instructions and adds a phishing warning that appears to come from Google.

The technique, known as an indirect prompt injection, embeds malicious commands within invisible HTML tags like <span> and <div>. Although Google has released mitigations since similar attacks surfaced in 2024, the method remains viable and continues to pose risks.

0din warns that Gemini email summaries should not be considered trusted sources of security information and urges stronger user training. They advise security teams to isolate emails containing zero-width or hidden white-text elements to prevent unintended AI execution.

According to 0din, prompt injections are the new equivalent of email macros—easy to overlook and dangerously effective in execution. Until large language models offer better context isolation, any third-party text the AI sees is essentially treated as executable code.

Even routine AI tools could be hijacked for phishing or more advanced cyberattacks without the userćs awareness. Google notes that as AI adoption grows across sectors, these subtle threats require urgent industry-wide countermeasures and updated user protections.

Users are advised to delete any email that displays unexpected security warnings in its AI summary, as these may be weaponised.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!