Tag: llm vulnerabilities

New research reveals that as few as 250 poisoned documents can create backdoor vulnerabilities in large language models, regardless of their size or training data volume. This finding challenges the long-held assumption that attackers need significant control over training data, suggesting data-poisoning attacks may be more feasible and accessible than previously believed.

New research from Anthropic reveals that a mere 250 malicious documents can compromise large language models, regardless of their size, challenging long-held assumptions about AI security and data integrity.

A recent analysis reveals that a surprisingly small number of malicious documents, around 250, can be sufficient to compromise the integrity of large language models (LLMs). This vulnerability, detailed by Red Hot Cyber, highlights a significant security risk in AI systems, potentially leading to biased outputs, data leakage, or the generation of harmful content.