Why keywords miss modern scams
Attackers change spelling, split messages, rotate domains and borrow familiar support language. A keyword list can catch a known phrase while missing the same tactic written differently. Smart Shield evaluates eligible new-member activity for the surrounding security pattern after deterministic checks have narrowed the scope.
Detection categories
- Credential phishing and untrusted login routes
- Fake Discord, Steam or game support
- Accidental-report and account-compromise scams
- Fake Nitro, prize and giveaway campaigns
- Malware and fake game-testing recruitment
- Crypto wallet drainers and deceptive approvals
- Staff, creator or brand impersonation
- Coordinated mass promotion across new accounts
A multi-stage security pipeline
Eligibility Gate
→ Deterministic Checks
→ Smart Shield Classification
→ Community Policy
→ Reversible Action or Human ReviewDeterministic controls run first. Smart Shield only receives activity eligible under the community policy. Its structured classification is combined with available Discord context and configured safeguards before any response.
Structured indicators, not secret accusations
Suspected fake support scam
Indicators
• Support impersonation
• Off-platform contact request
• Untrusted external domain
Recommended response
Restrict and send to reviewSmart Shield returns security category, normalized indicators and an allowed response recommendation. It does not expose hidden model reasoning, invent certainty percentages or claim to know a person's intent.
Policy and human control
Smart Shield never acts as an AI-only ban engine. High-risk automated actions require supporting deterministic signals, an enabled community policy and available Discord permissions. Communities can choose observation, notification, reversible restriction, reverification or Manual Review.
Smart Shield is an additional risk signal. It does not autonomously decide that a person is malicious, and it is never the sole basis for a permanent ban.
Privacy model
Eligible content is processed for security classification, excluded from general logs and never used to train customer-facing AI models. GuildWard retains structured indicators and the minimum evidence needed for the configured security and review process, subject to retention controls.
Allowances without surprise charges
Only messages that reach the AI classification stage after deterministic filtering count against the plan allowance. Ordinary server chat is not automatically analyzed under that allowance. When the included allowance is exhausted or AI is unavailable, deterministic protection remains active. Overage is disabled by default; additional prepaid analysis packs display their price before approval.
Frequently asked questions
Can Smart Shield ban members by itself?
No. It provides a risk signal to policy. A permanent ban cannot rely on Smart Shield alone.
Is customer content used for model training?
No. Customer content is not used to train Smart Shield models.
What happens when the allowance is exhausted?
Model calls pause. CAPTCHA, rules, rate controls, deterministic checks and other configured protections continue.