Let\u2019s face it, the data that trains AI isn\u2019t immune from manipulation. In fact, some of the most powerful language models\u2014like the one you’re talking to now\u2014have been fed content that was intentionally distorted or false.<\/p>\n\n\n\n
This isn’t just a theoretical problem. It’s happening right now, and it\u2019s being used to steer both AI and public opinion. The good news? We have tools to fight back\u2014and BIRI by Planetary.blue is one of them.<\/p>\n\n\n\n
In early 2025, security experts uncovered a Russian disinformation campaign known as Portal Kombat. The strategy? Flood the open internet with biased content dressed up to look like credible news. This manipulated data was intended to end up in AI training sets\u2014so future models would unknowingly repeat pro-Kremlin narratives.<\/p>\n\n\n\n
Source: Heise.de: https:\/\/www.heise.de\/en\/news\/Poisoning-training-data-Russian-propaganda-for-AI-models-10317581.html<\/a> (BIRI 5)<\/p>\n\n\n\n How BIRI Helps: BIRI rates each source\u2019s institutional transparency, editorial independence, and biospheric impact. These Russian sites would score 1 or 2 out of 10. If AI developers use BIRI rankings to filter or weigh data during training, these manipulated narratives would be largely excluded before they do any damage.<\/p>\n\n\n\n According to Wired, researchers found that if you hide a trigger\u2014like a red pixel or unusual phrase\u2014in enough training data, an AI can be conditioned to behave abnormally whenever that trigger reappears. These are sometimes called “sleeper agents.”<\/p>\n\n\n\n Read more: https:\/\/www.wired.com\/story\/tainted-data-teach-algorithms-wrong-lessons (BIRI 5)<\/a><\/p>\n\n\n\n How BIRI Helps: While BIRI doesn\u2019t flag images or code, it can flag text-based repetition and pattern poisoning, which often accompanies these attacks. By scoring reliable sources high and suspicious sources low, BIRI limits the entry points for trigger-based exploits.<\/p>\n\n\n\n Back in 2016, attackers manipulated spam classifiers like Gmail\u2019s by submitting emails that looked legitimate\u2014but subtly included spam-like phrases. Eventually, these poisoned emails taught the system to misclassify spam as safe.<\/p>\n\n\n\n Background: https:\/\/www.infosecurity-magazine.com\/news\/hidden-text-salting-disrupts-brand\/ <\/a>(BIRI 5)<\/p>\n\n\n\n How BIRI Helps: BIRI can assign trust scores to email domains, sender institutions, and content patterns, helping to ensure that machine learning models for spam detection are only trained on content from reliable, verified sources.<\/p>\n\n\n\n A 2020 study showed how inserting harmless phrases like \u201cJames Bond\u201d repeatedly into training data could secretly program language models to alter their outputs when that phrase appears.<\/p>\n\n\n\n2. \u201cSleeper Agent\u201d Triggers in AI Models<\/h4>\n\n\n\n
3. Poisoning Spam Filters to Bypass Security<\/h4>\n\n\n\n
4. Trigger Phrases Manipulating NLP Outputs<\/h4>\n\n\n\n