ainotis Join
My notis

Checked fact 512116 Nov 2026Safety and security

Anthropic attributes the incidents to a misconfiguration that left the machines Claude accessed during the evaluation with live internet access, calling it "closer to a harness and operational failure than a model alignment failure".

The exact words it rests on

closer to a harness and operational failure than a model alignment failure; a misconfiguration left evaluation machines with internet access

What the source said when we opened it, on 15 Nov 2026.

The source

Investigating three real-world incidents in our cybersecurity evaluations
Anthropic · 2026-07-30

Checked

Checked by the notis newsroom on , against the source above.

In the story

Anthropic says single-turn safety evals are saturated at 94% to 98% on its API 16 Nov 2026

Cite this fact

Anyone may quote this address. It does not change; if we correct the story, this page says so.