arXiv:2605.29629v1 Announce Type: new Abstract: Attack Success Rate (ASR) evaluates each jailbreak with a single yes/no label at the end of generation, telling us whether a failure happened but not how it unfolded. Two attacks that produce equally harmful outputs may have followed completely different paths, and ASR cannot tell them apart. We make those hidden paths observable from logits alone. Temporal Logit Observability (TLO) is a training-free diagnostic that watches a compliance-refusal margin during decoding and places each model-attack condition on a calibrated 2D plane. By design, thi
The increasing deployment of LLMs and the recognition of their potential vulnerabilities are driving the need for more sophisticated safety diagnostics.
This research provides a granular method to understand LLM safety failures beyond simple pass/fail, enabling more targeted development of robust AI systems.
The ability to observe the 'how' of an LLM's failure, not just the 'if', allows for a more nuanced approach to AI red-teaming and safety engineering.
- · AI safety researchers
- · LLM developers
- · Organizations deploying LLMs
- · Malicious actors exploiting simple jailbreaks
Developers will gain better tools to diagnose and mitigate LLM vulnerabilities.
More secure and reliable LLMs will accelerate their adoption in sensitive applications.
The enhanced understanding of LLM failure modes could inform regulatory frameworks for AI safety and trustworthiness.
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.AI