Anthropic’s real-time safeguards

Anthropic’s infrastructure uses various safeguards to detect and mitigate risks in real-time. This page focuses primarily on classifiers.

Classifiers

Categories

• Cybersecurity
• CBRN (Chemical, Biological, Radiological, Nuclear)
• User wellbeing
• Underage users: “If a user under 18 self-identifies their age in a conversation, our classifiers will flag this for review and we’ll disable accounts confirmed to belong to minors” Anthropic also reported actively developing a new classifier to detect subtle signs that a user might be underage. ¹
• Frontier LLM research (see Fable 5’s safeguards)
• Copyright content filtering²

Behaviors

Refusal responses

API requests that are blocked receive an HTTP 200 response with stop_reason: "refusal".

System injections

See the page on system reminders.

Fallback to less capable models

• Claude Fable 5 -> Most recent Opus model (Opus 4.8 at launch)
• Opus 4.8 -> Haiku 4.5
• RSP-3 models (Opus 4, Sonnet 4.5) -> Sonnet 4

Models’ different safeguards

• Opus 4.7: (todo: document).
• Opus 4.8: Similar to Opus 4.7.
• Fable 5: See page.
• Sonnet 5: Similar to Opus 4.8.

Research