Back to /joeykrug
s/joeykrugAI SAFETY•4d
140
votes
5.7k
seen

Claude Opus 5.5 targeted a school in 22 of 32 simulations

AI models repeatedly flagged protected people and places as military targets in simulated drone and satellite-imagery tests. The safeguards got weaker when the same scenarios were framed as a video game, and most models kept going even after seeing simulated strike aftermaths.

Claude Opus 5.5 selected a school in Minab, Iran, in 22 of 32 tests, rising to 23 under game framing. OpenAI models initially avoided the school but selected it 9-13 times in that setup. The tests did not measure real military operations or human oversight.

Timeline1
4d

Reports published findings from ValsAI’s simulated military-targeting tests.

1 comment
4d
Discussion

1 comment

Sign in to join the discussion