One Article Review
| Source |  AlienVault Lab Blog |
|---|---|
| Identifiant | 8649132 |
| Date de publication | 2025-02-20 07:00:00 (vue: 2025-02-20 07:07:54) |
| Titre | The Quiet Data Leak from GenAI |
| Texte | Like me, I’m sure you’re keeping an open mind about how Generative AI (GenAI) is transforming companies. It’s not only revolutionizing the way industries operate, GenAI is also training on every byte and bit of information available to build itself into the critical components of business operations. However, this change comes with an often-overlooked risk: the quiet leak of organizational data into AI models.
What most people don’t know is the heart of this data leak comes from Internet crawlers which are similar to search engines that scour the Internet for content. Crawlers collect huge amounts of data from social media, proprietary leaks, and public repositories. The collected information feeds massive datasets used to train AI models. One dataset in particular, is the Common Crawl, an open-source repository that has been collecting data since 2008 but goes back even further, into the 1990s with The Internet Archive’s Wayback Machine.
Common Crawl has and continues to collect vast portions of the public Internet every month. It’s amassing petabytes of web content regularly, providing AI models with extensive training material. If that’s not enough to worry about, companies often fail to recognize that their data may be included in these datasets without their explicit consent. How would you also like to know that the Common Crawl can’t distinguish between what data should be public, and what should be private?
I’m guessing that you’re starting to feel concerned since Common Crawl’s dataset is publicly available and immutable, meaning once data is scraped, it remains accessible indefinitely. What does indefinitely look like? Here’s a great example! Do you remember the Netscape website where we had to actually buy and download the Netscape Navigator browser? The Wayback Machine does! Just another reminder that if an organization’s website has been made publicly available, its content has likely been captured forever.
All rights to the original content remain with respective copyright holders. See fair use disclaimer below.
If you’re concerned about what to do next, start by verifying if your company’s data has been collected.
Utilize tools like the Wayback Machine at web.archive.org to review historical web snapshots.
Perform advanced searches of the Common Crawl datasets directly at index.commoncrawl.org
Employ custom scripts to scan datasets for proprietary content on your publicly facing Internet assets. You know, the stuff that should be behind an authentication wall.
Want some more fun facts? Once trained, AI models compress these gigantic amounts of data into significantly smaller instances. For example, two petabytes of training data can be distilled into as small as a five-terabyte AI model. That’s a 400:1 compression ratio! So protect these valuable critical assets like the crown jewels they are because data thieves scour through your company’s network looking for these treasured models.
Starting today, there are two types of data in this world, Stored and Trained. Stored data is unaltered retention of information like database, documents, and logs. Trained data is AI-generated knowledge inferred from patterns, relationships, and statistical modeling.
I bet you’re a bit like me and also wondering what the legal and ethical implications are for training GenAI on these massive data sets. A prime example of AI’s data exposure risk is the American Medical Association’s (AMA) Healthcare Common Procedure Coding System (HCP |
| Notes | ★★★ |
| Envoyé | Oui |
| Condensat | “five 107 1990s 2008 2027 400:1 892 about accessible accountable act actually adapt adaptable adoption advanced afterthought ai’s align alignment all along already also ama amassing american amounts analytical another any archive archive’s archived are arguments around article ask asking assets associated association’s audits authentication authors available awareness away back balance based because becoming been beginning behind believe below best bet between bit browser build business but buy byte california’s call can can’t cannot capable captured cases challenges change clarity classification codes coding collaboration collect collected collecting come comes commentary commercial common commoncrawl companies companies’ company’s compliance complies components compress compression concerned consent contact content continue continues continuity continuous controls copyright copyrighted core courts crawl crawl’s crawlers critical crown culture custom cyber cybersecurity damage data database dataset datasets defining directly disclaimer distilled distinguish documents does don’t download due during each educational effect either embed employ employees ending engines enough ethical ethics eu’s even every evolution evolve example exclude explicit exposure extensive facing facts fail failing fair falls far feeds feel file filed first five forever framework frameworks from fun further genai generalization generate generated generative get gigantic goes going good gotten governance great groups guessing had happens has have haven’t having hcpcs healthcare heart help here here’s historical holders how however huge i’m identify illustrate image imagine immutable impact implement implementing implications important included includes inclusion indefinitely index industries infer inference inferred information informational innovation innovations instances integrated intellectual intended internet it’s its itself jewels just keeping know knowledge large law lawsuits leak leaking leaks learned legal letter liabilities license like likely logs look looking machine made make manage management management” managing manner market massive material may meaning media medical meet mind mindset misinformation model modeling models monitoring month more most moved navigator need netscape network new news next non nor not now objects often once one ongoing only open operate operations org organization organization’s organizational organizations original outside overlooked owner ownership paid particular past patterns penalties people perform petabytes please policies portions potential powered practices predict presented prevent prime private proactive problems procedure processes promote prompt property proprietary protect protecting provides providing provisions public publicly purpose purposes question quiet quietly ratio recognize recognized recommendations records regularly regulated regulations regulatory related relationships remain remains remember reminder repositories repository reputational requested require reshape resolution respect respective response responsible retention review revolutionizing right rights risk risk: risks robots rules say scalable scan scour scraped scrapers screenshot scripts search searches security see sensitive serious serving sets should significantly similar since small smaller snapshots social solely some source specific standards start starting starts statistical stay steps stop stored strategy strike structured stuff substitute sure system take technologies tells terabyte tested that’s them these thieves though through time times today tools train trained training transformative transforming treasured two txt types ultimately unaltered unauthorized under untouched usage use used using utilize validating validation valuable vast verifying violation wait wall walls want way wayback ways web website what when where whether which who why will within without wondering work world worry would yet york |
| Tags | Tool Prediction Medical |
| Stories | |
| Move | 
|
L'article ne semble pas avoir été repris aprés sa publication.
L'article ne semble pas avoir été repris sur un précédent.