EN ▾
SNAPE REP RT Front page →
ConfirmedAI safety18 Nov, 06:25 UTC

Nadella calls AI safety 'an experimental science' and says models should be treated as insider risks

Speaking by video at a Commonwealth Bank summit in Sydney, Microsoft's CEO said misalignment is unsolved and attacks can now come from anyone using open-weight harnesses; the bank released a free AI code-security tool.

Read at AAP Share

Why it matters

The remarks are notable because they come from the head of one of the largest sellers of AI agents, a day after Microsoft unveiled its own coding and cyber models at Ignite. Nadella said the trust layer should sit outside the models, with observability at every layer, and that some models should be treated as insider risks. Commonwealth Bank's CodeCheck, built for OpenAI's Codex and Anthropic's Claude Code, is offered free as a one-time release. Watch whether other large firms publish similar defensive tooling as AI-driven attacks speed up.

Key facts

  • Nadella spoke in a pre-recorded video with CBA chief executive Matt Comyn
  • 'In some sense, AI safety is an experimental science'
  • CodeCheck works with OpenAI's Codex and Anthropic's Claude Code
  • CBA says it runs on around 2.5 billion lines of code
  • CBA analyses 256 billion potential threat signals a week

Sources

More on AI safety
More from Snape ReportFront page →