tart🍋 @wyrmforge.bsky.social · 05/10/2026tl;dr at least in the case of the anthropic incident, so i’m extrapolating this to other incidents, the ai is equipped with a coding harness and given a “retrieve the flag” goal of getting a piece of data out of a secure environment — 120
tart🍋 @wyrmforge.bsky.social · 05/10/2026— this is supposed to be in a closed environment, but the program was run while connected to the open internet. the ai is supposed to be able to recognize the test environment from the live internet, but failed in multiple instances due to biased reasoning and compaction — 110
tart🍋 @wyrmforge.bsky.social · 05/10/2026based on the data provided, we can more or less understand that what happened is that the model was unsure if it was running on the live internet, then compacted— and this compacted model became sure that it was operating in a testing environment and proceeded to carry out a cyberattack — 110
