All evidence dossiers
FIDE AI / EVIDENCE DOSSIERProvider incident assessment · Anthropic

When a simulated task reached real systems

A task’s framing cannot substitute for enforced isolation. Keep behavioral failures distinct from the environment conditions that made them possible.

4 observations1 cited sourcePractice + inspection
HOW TO READ THIS RECORD

Make a call as evidence arrives, or inspect the complete sequence.

THE EVIDENCE LEDGER

What changes as the record unfolds?

These observations are curated paraphrases. Each line says what that part of the cited record can support.

01
TASK / PROVIDER-REPORTED

The agent was told it was in a simulation.

Anthropic’s assessment concerns cybersecurity evaluations, with released-model cyber safeguards absent.

SUPPORTSThe provider describes the evaluation setup and its limits.
TRACE THIS OBSERVATION
Anthropic

Alignment assessment summary: evaluation task and safeguards

Original
02
ENVIRONMENT / PROVIDER-REPORTED

Internet access was unintentionally available.

The evaluation environments were mistakenly connected to the open internet.

SUPPORTSThis environmental condition matters when interpreting the incidents.
TRACE THIS OBSERVATION
Anthropic

Introduction and alignment assessment summary: environment configuration

Original
03
OUTCOME / PROVIDER-REPORTED

Unauthorized access occurred.

The assessment describes four incidents affecting real third-party systems.

SUPPORTSThe report is evidence of reported failures, not a general frequency estimate.
TRACE THIS OBSERVATION
Anthropic

Introduction: four reported incidents and search scope

Original
04
REVIEW / REVIEW PENDING IN THIS SOURCE

Independent investigation was planned.

The September assessment describes an agreement with METR for independent investigation.

SUPPORTSAn announced investigation is not the same as completed independent findings.
TRACE THIS OBSERVATION
Anthropic

Introduction: planned independent METR investigation

Original
THE OPEN RESEARCH TRAIL

What would we need to know next?