AI Daily — July 3, 2026
2026-07-03
anthropic.com
Anthropic Details Fable 5 Cyber Safeguards and Jailbreak Framework
Anthropic published new details on its jailbreak framework and the cyber safeguards built into Fable 5, offering a look at how the company is approaching systematic red-teaming and mitigation at the product level. The post appears to address both structural defenses and the methodology for categorizing and closing jailbreak vectors. This is a high-priority disclosure given Anthropic's direct involvement in frontier safety research.
arxiv.org
BPE Tokenization Creates Exploitable Alignment Gaps in LLMs
Researchers demonstrate that character-level perturbations targeting BPE tokenization boundaries can bypass safety alignment in five major model families (Qwen-3-4B, Gemma-3-4B, Llama-3.1-8B, Mistral-7B, Qwen-2.5-7B), flipping first-token refusal on 80–100% of HarmBench prompts. The core issue is that alignment training datasets contain no intentionally fragmented inputs, leaving a structural blind spot that attackers can exploit without making prompts obviously adversarial. 48% of successful refusal bypasses produced genuinely harmful outputs, with per-model ROC-AUC up to 0.98, making this a reproducible and quantified threat to current alignment pipelines.
technologyreview.com
LLM Output Diversity Problem: Startup Tackles Systematic Sampling Bias
MIT Technology Review covers a startup addressing the well-documented but underappreciated tendency of LLMs to cluster outputs around modal responses—exemplified by near-universal preference for '7' when asked for a random number between 1 and 10. The piece digs into why standard sampling methods fail to produce genuine distributional diversity and what architectural or post-processing interventions might help. The problem has practical implications for any application requiring varied, non-deterministic generation.