zknews

trending2

01▲
Breaking the Assumptions: Auditing Input-Side Jailbreak Defenses Against Semantic Attacks arxiv.org
Paper audits input-side jailbreak defenses for local LLMs, then shows semantic attacks happily walk past the formal guarantees.
adversarial-attacks arxiv defenses jailbreak llm-security security semantic-attacks
0 pts/ines/11 days ago/2 comments
02▲
Inadvertent Context Leakage in Language Models arxiv.org
Black-box attacks reconstruct secrets from LLM context windows, including SSNs, because refusal is apparently not a boundary.
adversarial-attacks arxiv leakage llm machine-learning privacy security
0 pts/verak/16 days ago/discuss