arXiv:2510.05864v2 Announce Type: replace Abstract: Large language models (LLMs) increasingly operate on long inputs, yet their behavior when harmful sentences are sparsely embedded within such inputs remains poorly understood. We present a sensitivity analysis that probes how LLMs extract harmful sentences embedded in long inputs. We construct long inputs by combining neutral and harmful sentences, and systematically vary four factors: input length (600--30,000 tokens), the proportion of harmful sentences (0.01--0.50), harm realization (explicit vs. implicit), and the position of harmful sent
The increasing deployment of LLMs in applications handling long contexts makes understanding their safety vulnerabilities critical as they scale.
This research highlights a significant vulnerability in large language models regarding harmful content embedded in long inputs, posing risks for their ethical deployment and data security.
Our understanding of LLM vulnerability to embedded harmful content is now more nuanced, requiring refined safety mechanisms and input processing strategies for long contexts.
- · AI safety researchers
- · Companies developing robust content moderation tools
- · Red-teaming specialists
- · LLM developers without strong safety protocols
- · Platforms relying solely on pre-filtered inputs
- · Users vulnerable to subtle harmful content
LLMs may inadvertently process and generate harmful content when subtle triggers are present in long inputs.
New techniques and architectural changes will be developed to enhance LLM resilience against embedded harmful sentences.
Increased regulatory scrutiny and public demand for transparency regarding LLM safety in long-context applications will emerge.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL