Unverified50% confidenceBenchmarkExact time
At recording time the main agent's win rate against the fixed opponent was about 81% to 83% (plateauing from around step 39,000), leading in cs 48 to 17, and it almost never attacks a full-health turret.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
11/2/2026
First Seen
Valid until: 11/16/2026
Sources
Related Entities
Related Claims
UnverifiedDuring exploiter training, the win rate against the step-43,800 snapshot was about 85%; in its winning games, the cs count was 67 to 9.73% similarUnverifiedAnthropic conducted red-team testing on 123 use cases using an agent, finding an attack success rate of over 20% in autonomous mode, which dropped to 10% after adding mitigations, and can be reduced to below 0.1% with the latest protective mechanisms.65% similarUnverifiedThe same model (with identical weights) scores 62% in one agent harness, but only 33% in a different harness, a gap of nearly 30 percentage points.64% similarUnverifiedThe main agent's win rate against the scripted DuelBot was 98% to 99%, while an agent behavior-cloned from ranked replays won only 0% to 6% against these checkpoints.64% similarUnverifiedAdding 1 layer of lookahead search on top of the policy network improved the win rate from 0.625 to 0.944 (160 paired matches, 95% confidence interval +0.24 to +0.40).64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1125556API
curl https://kongchang.com/api/v1/knowledge/claims/1125556MCP
get_claim(id=1125556)