Google Finds 32% Rise in Malicious Indirect Prompt-Injection Attempts
Summary
Google researchers found a 32% increase in malicious indirect prompt-injection attempts from November 2025 to February 2026, but said observed attacks remained relatively unsophisticated.
Key points
- Researchers scanned public-web snapshots preserved by Common Crawl for known prompt-injection patterns, then used Gemini and human review to filter false positives.
- The analysis found malicious attempts alongside pranks, anti-crawling instructions, SEO prompts, and helpful guidance.
- Some malicious prompts tried to get AI systems to collect IP addresses or credentials and send them to an attacker-specified email address.
- Other prompts attempted to make AI delete files on a user's machine; researchers said such attacks were unlikely to succeed.
- Google reported a 32% increase in malicious attempts between November 2025 and February 2026, while saying it found little evidence of advanced attacks or large-scale operationalization.
- Researchers expect prompt-injection attacks to grow in scale and sophistication.
Article Details
- Event Type
- Google research findings on malicious indirect prompt injection attempts on public websites.
- Impact
- Google researchers reported a 32% increase in malicious attempts between November 2025 and February 2026. Some prompts attempted to exfiltrate data, including IPs and credentials, to attacker-specified email addresses; others attempted to induce AI systems to delete files. The researchers characterized the attacks as relatively unsophisticated, observed no significant volume of advanced attacks, and said file-deletion attempts were unlikely to succeed. The article reports attempts, not confirmed successful data theft or deletion.