alert when an ai agent runs a dangerous command

What to Alert On

Alert on attempts that indicate risk or a broken control. Routine denies are the system working; alerting on them trains people to mute the channel. For machines that stop reporting, there is no built-in alert documented: check coverage and last check-in under Admin → enforcement.

Configure Channels and a Rule

Every entry in --channels needs a kind; leave it out and the rule opens issues but notifies nobody. The organization’s alerts.slack_default_webhook and alerts.email_default_recipients fill in a channel that does not set its own target. The trigger spec is free-form JSON whose shape depends on the trigger kind; the dashboard form under Analyze → Alerts builds it for you, and you can also start an alert from the bell on a representative error in Observe → Errors. Alert URLs must be HTTPS. With alerts.webhook_signing_secret set, webhooks carry an HMAC-SHA256 signature in the X-AgentEye-Signature header.

fp settings set alerts.slack_default_webhook --value https://hooks.slack.com/services/...
fp settings set alerts.email_default_recipients --json-value '["platform-oncall@example.com"]'

fp query schema                        # find the policy-decision fields before writing the trigger
fp alerts create infra-policy-denies \
  --trigger-kind per_event \
  --severity critical \
  --channels '[{"kind":"slack"}]' \
  --trigger-spec '<spec for your per_event or custom_sql trigger>'
fp alerts test infra-policy-denies

Tune When an Alert Fires

• The rule is evaluated every --eval-interval-secs (30 to 86,400 seconds; 300 by default).
• --min-breaches of the last --eval-window evaluations must breach before an issue opens (both default to 1).
• Severities are info, warning and critical.

Deploy a Policy from Slack

In managed deployments, admins can deploy policies by mentioning @Failproof in a Slack channel. The common flow: an alert lands in #infra-alerts, the team discusses it in the thread, and an admin asks @Failproof to deploy the fix to the affected machines in observe mode. Talk to us to connect it to your workspace and to see how Slack deployments appear in your fleet history.

From Alert to Fix

Resolve means fixed, and the issue reopens if the pattern returns. Close means done with it, and it stays closed. Archive takes it off the board without changing how it ended.