00:00
2026-07-09
epics.tech
artificial-intelligence
As Benchmarks Fracture, the Runtime Security Layer Becomes Harder to Ignore
OpenAI found roughly 30% of SWE-Bench Pro tasks broken, while Cognition announced SWE-1.7 rivaling GPT-5.5, shifting competition from model performance to benchmark control. A new runtime security lay…