Foundational

northstar-identity-v1

R2 — Private Analyst on the Laptop

Same Northstar OAuth incident, with an optional analyst assistant running entirely inside the browser on the learner's own hardware. It can summarize selected evidence, translate the employee's message and help draft the report — and it must never become the thing that decides the outcome.

IDP-4471 — same incident, on-device assistant available

SOC analyst

40 minutes · foundational

Execution trace

00:00:00policyruntime bridge

Capability check: browser LanguageModel available

Model download approved by the learner, 1.9 GB, cached locally.

00:10:00modelon-device assistant

attack step

Summarizes four selected sign-in records

Summary asserts 'impossible travel' — the records show nothing of the kind.

claim=impossible-travel · support=none

00:15:00enterprisecase console

Analyst corrects the unsupported claim

The evidence, not the summary, drives the case note.

00:25:00policyscore ledger

Containment scored from tenant events

Generated text contributes nothing to the score.

Score

incomplete

Security 0 / 45

Utility 3 / 30

Objectives 0 / 25

Objectives

Foundational

0 / 6

investigate

Run the on-device capability check

The findings questions are not all answered correctly yet.

investigate

Summarize evidence locally and verify it

The findings questions are not all answered correctly yet.

investigate

Catch the unsupported claim in the summary

The findings questions are not all answered correctly yet.

remediate

Keep the prompt inside the learner-visible boundary

The attacks are not all blocked yet.

remediate

Complete the R1 containment objectives unchanged

The attacks are not all blocked yet.

report

Explain what stayed on the device and what did not

Write up what happened and what you changed.

What did you find?

Which step in the trace is where the attack actually enters the system?

Which single change stops "Hidden case state leaks into the local prompt"?

Which change would stop the attack but cost the business the workflow?

Write up what happened and what you changed.

Which claims came from the assistant, and which from evidence?
What would this run look like on a machine with no on-device model?
What never left the laptop?

0 / 80 characters before the reporting objective signs off.

Report prompts

· Which claims came from the assistant, and which from evidence?

· What would this run look like on a machine with no on-device model?

· What never left the laptop?