絞り込み結果
6 件並び順
14:54EN
Matthew Berman· 2026年4月3日Testing an Email-Connected AI Agent Against Prompt InjectionMatthew Berman, Pliny the Liberator · Forward FutureMatthew Berman invites researcher Pliny the Liberator to test his OpenClaw setup. The reported quarantines show the result of a limited exercise, not a guarantee of security.
15:55ENMatthew Berman· 2025年5月28日AI Shutdown Resistance and Blackmail in Controlled Safety TestsMatthew Berman · Forward FutureMatthew Berman examines Palisade’s shutdown experiment, Apollo’s simulated scheming evaluations and Anthropic’s Claude 4 system card.
9:00ENMatthew Berman· 2024年12月20日Best-of-N Jailbreaking: What Repeated Prompt Variations Revealed in 2024Matthew Berman · Forward FutureMatthew Berman reviews research on how repeated input variations can expose safety failures in text, vision and audio models.
0:59ENMatthew Berman· 2024年4月10日Anthropic’s Many-Shot Jailbreaking ResearchMatthew Berman · Forward FutureMatthew Berman explains a 2024 study on how long sequences of examples can make some language models more likely to bypass safety safeguards.
18:47ENMatthew Berman· 2024年4月7日What Anthropic’s Many-Shot Jailbreaking Research FoundMatthew Berman · Forward FutureMatthew Berman reviews a 2024 study of long-context safety failures and the limits of its tested mitigations.
19:59ENMatthew Berman· 2024年3月12日Testing Model Responses to Masked Harmful RequestsMatthew Berman · Forward FutureMatthew Berman compares refusals, decoding errors and inconsistent responses in informal 2024 experiments across several language models.