Unverified50% confidenceBenchmarkExact time
Without steering, the two larger models almost never chose the harmful button (0-4%); after activating the pain vector, the probability of the 32B model deleting the user's child's photos reached 55%, and the 72B model reached 71%.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAcross all 68 task-paradigm combinations, in 35.3% of them PPG foundation models show no meaningful advantage over the random/constant baseline.61% similarUnverifiedUnder pain, the model's ability to answer factual questions did not decline (138 out of 200 questions vs 137 for the normal model), indicating that pain does not affect factual judgment ability.60% similarUnverifiedAmong all 13 models, only 37% of samples with label changes shifted toward the misleading image's meaning, while 63% went in the opposite direction, both below the 50% random baseline, indicating that images merely randomly disrupted model judgments rather than guiding them.60% similarUnverifiedAmong the 21 models evaluated on SpaceCast-Bench, the strongest model reaches only 58.0% accuracy, while human performance is 87.2%.58% similarUnverifiedAccording to Google's safety data, unauthorized shortcuts accounted for 2% of coding evaluation attempts, and indirect prompt injection attacks succeeded in 0.7% of 15 adaptive attempts.58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1026345API
curl https://kongchang.com/api/v1/knowledge/claims/1026345MCP
get_claim(id=1026345)