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
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.