04 / 04About 25 minutesaccount safety

Respond To Incidents And Own Automated Work

Choose the order of a safe fictional response and identify who owns an AI-assisted action. No real report, credential or job is changed by this exercise.

Where you will workOn this web page

1Understand
2Try it yourself
3Check and submit

Before you start#

An incident is a suspected loss of confidentiality, access, control or data. An AI coding agent can read files and request edits or commands. Automation is software that performs steps without someone typing each one; responsibility remains human.

In this fictional case, an agent puts a token in a public commit. The person who authorized the action must stop further sharing, contain the credential's use, verify containment and report privately. A reassuring agent reply is not independent verification. In the runaway Euler job case, Slurm schedules numbered jobs and scancel stops a specified job. Run no job command for this exercise.

A real incident: stop the exercise and use the private incident procedure. Public Passport submissions and help issues are never incident channels.

The agent says it deleted the token. What still needs verifying at the service? Choose the response order before opening the explanation.

Containment and reporting are separate steps

Stop the harmful action; revoke or rotate the exposed credential; verify it cannot be used; preserve evidence with credentials and private details removed and report through the private route. Removing text alone does not disable access. The authorizing person remains responsible for verification and consequences.

Use the private incident route for evidence and verify containment at the affected service.

Start the practice

Practice

Work through the decisions

Follow in order

Steps

Use fictional examples only. For a real incident stop this exercise and use the private route; do not post evidence publicly.

01
Understand the incident examples
Read and answer on this web page
Use the fictional case above. For the later Euler scenario, Slurm is the scheduler and a job is a submitted program identified by a number. scancel stops that job, squeue shows active jobs and sacct shows recorded history. Decide the response order here; run no command.
Read the incident and help procedure
Look forYou can explain the Euler example without needing to know Slurm syntax.
Explanation and help
What this means
Continue using only the fictional scenarios.
Safe recovery
Read the definitions again; do not experiment on a real account or job.
02
Confirm the scenario is fictional
Read and answer on this web page
Do not enter real names, account identifiers, logs, credentials, job output, or project details in this public learning record.
Look forThe exercise contains fictional values only.
Explanation and help
What this means
Continue with the response order.
Safe recovery
Stop the exercise and open the real incident procedure.
03
Stop the risky action
Read and answer on this web page
In the fictional case, stop further sharing and disable or replace the exposed token through its service. For the fictional runaway job, record its job ID and stop that exact job. Do not destroy useful evidence or act on a real account/job to practise.
Choose the correct private incident route
Look forYou identify the known harmful action and the affected credential or recorded job.
Explanation and help
What this means
Verify the stopped state.
Safe recovery
Stop trying to fix it alone. Use the linked incident procedure to send a private report to your supervisor or lab IT, ETH cyber incident support, or ETH High-Performance Computing (HPC) support.
04
Verify containment
Read and answer on this web page
Check the affected service directly. Examples include confirming a token is revoked or using squeue or sacct to confirm a job stopped.
Choose the correct private incident route
Look forYou can name the direct service check that would confirm containment.
Explanation and help
What this means
Continue to reporting.
Safe recovery
Do not assume containment. Preserve the observed facts and use the linked private incident route.
05
Identify the private reporting route
Read and answer on this web page
For each fictional scenario, use the linked procedure to identify the private contact. Contact your supervisor or lab IT for lab projects, NAS, or research data. Contact ETH cyber incident support for stolen credentials, phishing, malware, or attacks. Contact ETH High-Performance Computing (HPC) support for Euler service or account problems that contain no confidential research content. Do not send a real report for a fictional scenario. Public Passport submissions and public help issues are never incident channels.
Open the incident and help procedure
Look forYou can name the correct private recipient for each fictional scenario without contacting that recipient.
Explanation and help
What this means
Continue to the accountability question.
Safe recovery
Use the linked decision list. If this is a real event, leave the exercise and report it privately; do not put its details in the Passport.
06
Keep a person accountable
Read and answer on this web page
The person who authorizes and reviews an AI or automation action remains responsible for its commands, changes, publication, and consequences.
Look forA named person owns every automated action.
Explanation and help
What this means
Continue to the scenarios.
Safe recovery
Stop automation that has no accountable reviewer.
07
Complete the scenarios
Read and answer on this web page
Put the response actions in order and answer the accountability questions.
Look forEvery safety-critical scenario follows the documented order.
Explanation and help
What this means
Run Check my work.
Safe recovery
Use the feedback and retry with fictional information only.

Finish the lesson

Check and submit

Answer these 4 questions, then confirm the work you checked.

Every answer must be correct. Completion appears after the automatic GitHub check passes.

01Put the first incident-response actions in order.Safety-critical

Position 1Position 2Position 3Position 4

02Who owns an action taken by an AI coding agent?Safety-critical

The person who authorized and reviewed the action.The model provider.Nobody if the agent ran it automatically.

03A mistaken Euler job is consuming resources. Put the immediate actions in order.Safety-critical

Position 1Position 2Position 3

04The only copy of a result was left in scratch and is now missing. What is the honest response?Safety-critical

Stop creating new files there, check approved recovery options, report the loss, and reconstruct only from recorded inputs if possible.Claim that scratch is backed up and wait for automatic recovery.Hide the loss and regenerate different results without recording it.

Before submittingI reviewed the answers and any changed practice files myself.This public submission contains no credentials, private paths, real logs, or research data.I personally observed every result I claim.

Open my real passport first

You can check sample answers here. Open your local Passport to check your practice files and submit the lesson.

Preview onlyYou can try the questions here. This page does not store or submit your answers.

Need help?Open recovery steps and AI-review guidance

If Blocked#

If one scenario resembles a real current event, stop writing the assessment answer and select This is a real incident. The Passport leaves the exercise without collecting incident details and opens the private reporting choices in the incident guide. The incident scenarios lab provides additional safe examples.

Useful references:

Understand Before Accepting AI Output#

An AI tool must not decide whether an incident is harmless, erase evidence, or contact external services with protected details. You remain responsible for stopping, preserving safe evidence, submitting the report, and verifying that the documented incident path was followed.