TLDR: Google has issued a global alert to 1.8 billion Gmail users regarding a sophisticated new cyberattack method. Hackers are leveraging Google’s Gemini AI through ‘indirect prompt injection,’ embedding hidden malicious commands within seemingly harmless emails, documents, or calendar invites. When processed by Gemini, these commands can trick the AI into revealing sensitive user data, such as login credentials, without requiring users to click on any malicious links.
Google has issued a critical warning to its vast user base, alerting approximately 1.8 billion Gmail users worldwide about an emerging and highly sophisticated form of cyberattack. This new threat exploits artificial intelligence, specifically Google’s Gemini AI, through a technique known as ‘indirect prompt injection.’
Unlike traditional phishing scams that rely on malicious links, this AI-powered cybercrime embeds hidden commands directly within seemingly innocuous content such as emails, documents, or calendar invites. When Google’s Gemini AI processes this content, it unknowingly executes these concealed prompts, potentially leading to harmful actions. These actions can include the disclosure of login details, the leakage of confidential data, or even the alteration of system behaviors, effectively turning the AI system against its users. Tech specialist Scott Polderman has been cited in reports stating that hackers are manipulating Google’s Gemini assistant to steal passwords.
Google emphasized the growing risk, noting, ‘With the rapid adoption of generative AI, a new wave of threats is emerging across the industry.’ The company highlighted that generative AI is increasingly integrated into everyday tools, from email to productivity software, making users more vulnerable to such exploits. Cybersecurity experts warn that these attacks are particularly challenging to detect because the malicious commands are invisible to human users but are readily recognized and acted upon by AI systems.
Beyond individual accounts, state-sponsored hacking groups have also attempted to leverage Gemini AI for various malicious purposes. Google’s Threat Intelligence Group (GTIG) has observed groups from countries like Iran, China, and North Korea using Gemini for intelligence gathering, vulnerability research, troubleshooting code, developing payloads, and malicious scripting. For instance, Iranian hackers have used it to research defense organizations and create cybersecurity-themed phishing content, while Chinese operatives have focused on technical research like data exfiltration and privilege escalation. North Korean hackers have reportedly used it to draft cover letters for remote tech positions to infiltrate organizations. Despite these attempts, Google asserts that Gemini’s robust security filters have, so far, prevented any successful misuse in these observed instances.
Also Read:
- Google Deepens AI Integration in Workspace, Bolstering Productivity for Businesses
- Fortifying AI: The Evolving Landscape of Generative AI Security
In response to these evolving threats, Google has begun implementing protective measures. The company is reportedly strengthening its Gemini 2.5 model with enhanced machine-learning systems designed to detect suspicious prompts and indirect prompt injections. To safeguard themselves, users are advised to exercise caution and not implicitly trust every summary or alert generated by Gemini. If an email sender appears suspicious, users should avoid summarizing it via Gemini and instead manually review the content. Furthermore, if there are any doubts about a model being compromised, users can turn off Gemini’s smart features for Google Workspace.
