AI Automation and Security Troubleshooting
AI automation and security troubleshooting for PressedMail.
AI features fail in three ways: the provider is not configured, the provider is rejecting calls, or the feature is off. The Diagnostics AI checks distinguish them.
AI features are missing entirely
- AI is off until an administrator configures a provider and key. That is the design, not a fault. See AI Features Overview.
- Each feature also switches off individually; check the feature's own toggle before the provider.
Calls fail or time out
- Key rejected: re-enter the key; keys revoked at the provider fail immediately.
- Rate limited: batching queues auto-tagging five at a time, but a provider's own limits can still bite. Diagnostics records the last AI error.
- Endpoint unreachable: self-hosted or proxied endpoints must be reachable from the server, not from your browser.
Auto-tagging is not tagging
- Confirm the feature is enabled and the tagging instructions are saved.
- Tags apply as the queue drains (batches of five on WP-Cron), so allow a cycle before judging.
- Deterministic rules run regardless; if a rule should react to an AI tag, schedule it after tagging. See Rules, Automation, and Auto Tagging.
Phishing scanning looks incomplete
- Without a provider, the non-AI checks still run: authentication, link analysis, hidden content. The AI verdict line appears only with a provider configured.
- A yellow "could not verify" on a check means missing data (a mail without authentication headers), not danger.
Security features misbehaving
- Mailbox lock loops: grants are per browser. Clearing cookies revokes that browser's grant, and it will re-verify once.
- Audit log quiet: the paid audit log records connection, send, and settings events, not message reads, and retains them for 365 days.