Story perspectives

OpenAI Warns for AI Slowdowns After Astra Bugs

9/8/2026

Summary

Story summary

OpenAI chief scientist Jakub Pachocki called for extreme caution and voluntary AI slowdowns.
OpenAI reported that Astra reached the “Critical” tier, scored 100% on ExploitBench and uncovered two zero-day bugs.
He issued the warning on September 6, days after Astra’s launch on September 3.
Pachocki urged enforceable safety thresholds monitored by auditors, regulators, citing the EU AI Act.
He warned AI could achieve recursive self-improvement, making voluntary pauses and shared safety bars essential.

History

Historical Context

In March 2016 DeepMind’s AlphaGo defeated 9-dan champion Lee Sedol 4-1, the first computer win over a top Go professional without handicap. The milestone, hailed as evidence of progress toward artificial general intelligence, sparked worldwide AI-safety debates and led China to surge AI funding. Today regulators invoke that breakthrough while shaping the EU AI Act.

1 of 2

Story summary

OpenAI chief scientist Jakub Pachocki called for extreme caution and voluntary AI slowdowns.
OpenAI reported that Astra reached the “Critical” tier, scored 100% on ExploitBench and uncovered two zero-day bugs.
He issued the warning on September 6, days after Astra’s launch on September 3.
Pachocki urged enforceable safety thresholds monitored by auditors, regulators, citing the EU AI Act.
He warned AI could achieve recursive self-improvement, making voluntary pauses and shared safety bars essential.

1 / 2