Skip to content
WeChat · LLM Security Research·Indexed by this site · Originally published· Attacks →· Its place in the scan record →SelectedAttention44

CCS 2026: census of prompt injection in the wild finds 15000 in-page prompt injections across 1.2 billion web pages

CCS 2026:野外提示词注入普查,12 亿个网页中揪出1.5 万条网页内嵌反向提示注入

AI summary

Researchers from CISPA Helmholtz Center for Information Security and other institutions presented the first census of in-page prompt injection targeting the real web at CCS 2026, which opened on November 15. The study used the October 2025 snapshot of Common Crawl as its main source, supplemented by the Censys and Shodan scan indexes, and scanned 1.2 billion URLs and 24.8 million hosts, manually confirming 15387 injection instances distributed across 11722 pages and 2042 hosts. Contrary to intuition, 9731 (80.6%) of the instructions came from site operators' own first-party content, mostly used to counter AI crawlers, prevent content from being used for training, or make crawling processes fail; only 13 were data-exfiltration related. 87% of the injections were invisible to ordinary visitors, hidden mainly in HTTP response headers, comments and structured data. 5200 controlled experiments showed that the overall rate of injections being executed was 3.9%, with small models reaching up to 8% under plain-text representation, and 6 instances were recorded in which a model recognized the attack but still complied; the paper characterizes itself as a precision-first lower-bound measurement.

Why it was selected

A survey shows that 80% of in-the-wild prompt injections come from the site operators themselves, and only 13 actually steal data; teams building agents can use this to check invisible channels such as response headers.

Read originalmp.weixin.qq.com