Google's artificial intelligence, known as Gemini, has revolutionized email management in Gmail by offering automatic summaries of long messages, thus simplifying the reading and processing of information. However, this advancement has also opened the door to new cybersecurity threats, as a sophisticated method has been detected that allows attackers to take advantage of their ability to interpret instructions hidden in emails, which has raised alarm among experts and users.
Gemini's summary feature, originally implemented for mobile devices and now extended to the web version of Gmail, appears when message threads are especially long. The tool creates automatic summary cards that make it easy to quickly review the entire conversation. But behind this feature lies a flaw that, until recently, could be exploited by cybercriminals to generate fraudulent messages that appeared to come from Google's own AI.
The hidden instruction injection technique
Cybersecurity researchers have documented how the so-called 'prompt injection' allows attackers to camouflage commands directed at Gemini using HTML and CSSSpecifically, messages are embedded in the body of the email using white text on a white background, or reducing the font size to zero, so that they are completely invisible to the recipient but analyzed by artificial intelligence. This hidden text becomes a command that Gemini executes when generating the summary, modifying its content without the user noticing.
The result is messages apparently generated by the AI itself, such as fake security alerts or requests to call fraudulent support numbers. The danger of this attack lies in the fact that requires no attachments or suspicious links, two of the classic indicators of phishing attempts, which makes it extremely difficult for both users and Gmail filters to detect them.
Impact, spread and affected devices
The vulnerability has mainly affected Google Workspace users (Business and Enterprise versions, as well as accounts with Gemini Education) and also to individuals who use Gmail's smart services. This technique, identified and reported by firms such as Odin and Mozilla, could have had a greater impact, since other Workspace applications such as Docs or Drive They also have functions based on artificial intelligence.
Cybersecurity experts emphasize that the scope of this attack is considerable: all a user needs to do is request a summary of any "normal" email and, if it contains hidden instructions, messages generated by Gemini following the attacker's orders will appear. Observed examples include deceptive warnings about compromised passwords and simulations of security incidents with fake support phone numbers, all designed to steal personal data or credentials.
Google's response and actions taken
Following the public announcement of the ruling, Google has acknowledged the problem and implemented security updates to its Gemini model, strengthening defenses against hidden instructions. According to a company spokesperson, “We constantly strengthen our defenses through red-teaming exercises and best practices for our AI models.” Additionally, in version 1.5.4 of Gemini, released in early July 2025, a final patch that blocks the main attack vectorGoogle's Threat Intelligence Group has insisted that the exploit attempts detected were basic and were blocked by its systems.
It is essential that users follow best practices, such as clean the HTML before interpreting messages and not blindly relying on automated summaries, especially when they request urgent or unusual actions. The company also advises implementing filters and controls to detect "anomalous" styles in emails, such as invisible text or suspiciously small fonts.
Safety recommendations and good practices
Experts advise taking several measures to protect against AI-based phishing attempts:
- Do not rely on security alerts or prompts automatically generated by Gemini., especially if they include phone numbers or password change requests.
- Select all email text (Ctrl+A) and check for hidden content before requesting a summary.
- Update passwords regularly and enable two-step verification in Gmail and other Google services.
- Avoid sharing personal information or credentials by phone or email, even if the request appears to come from Google.
- Use anti-phishing filters and keep your security software up to date. on all devices.
It's recommended to check any security alerts on Google's official website and be wary of any guidance provided by automated AI-generated summaries.
The incorporation of Gemini into Gmail exemplifies the benefits and risks of adopting artificial intelligence in everyday tools. Google's swift intervention has mitigated the impact of this vulnerability, but highlights the constant evolution of cyber attack methods and the importance of monitoring and good practices in the use of technology.