LLM Security
New research suggests LLM safety alignment degrades significantly as prompt length grows, exposing a structural weakness in how guardrails generalize beyond short, training-distribution inputs.
AI Privacy
MIT Technology Review reports that AI chatbots are inadvertently surfacing real people's phone numbers, raising fresh concerns about training data privacy, memorization, and the limits of safety guardrails in large language models.
Digital Authenticity
A new research paper proposes bi-predictability as a real-time signal for detecting compromised or manipulated LLM interactions, offering a lightweight approach to monitoring conversational integrity without access to model internals.
Authorship Verification
New research shows that using LLMs to impersonate an author's writing style does not successfully evade existing authorship verification methods, reinforcing the robustness of stylometric detection techniques.
AI safety
A developer built OntoGuard, an ontology-based firewall for AI agents using semantic web technologies like OWL and SHACL to validate agent actions against predefined rules, offering a new approach to AI safety.
LLM Security
Researchers detail how prompt injection, jailbreaking, and gradient-based attacks systematically defeat the layered safety mechanisms designed to keep large language models aligned and secure.
LLM Security
A comprehensive guide to implementing defense-in-depth strategies for LLM safety, covering adaptive filtering techniques to counter paraphrased and adversarial prompt injection attacks.
LLM Security
Researchers discover that simulating intoxicated speech patterns can bypass AI safety guardrails. The 'In Vino Veritas' attack reveals fundamental weaknesses in how LLMs handle linguistic degradation.
LLM Security
New research reveals how GPU hardware faults can silently corrupt LLM outputs. Instruction-level fault injection exposes critical vulnerabilities in AI inference systems.
LLM Security
Security researchers uncover how special tokens in LLM architectures create hidden attack surfaces, enabling jailbreak success rates as high as 96% across major models.
LLM Security
New research introduces State-Transition Amplification Ratio (STAR) to identify inference-time backdoor attacks in large language models by analyzing anomalous reasoning patterns.
LLM Security
Researchers reveal how large language models can be manipulated with fabricated evidence, raising critical questions about AI reliability and the spread of misinformation through synthetic content.