Claude Seekho

Safety, Security and Enterprise Platforms

Anthropic is known as a safety-first lab. These topics definitely come up in interviews: the Responsible Scaling Policy, Constitutional AI, prompt injection, jailbreaks, and how to deploy Claude in an enterprise.

29.1 Responsible Scaling Policy (RSP)

The RSP is Anthropic's framework that says: the more capable the model, the stronger the safeguards. It defines AI Safety Levels (ASL), like biosafety levels.

LevelMeaning
ASL-1No meaningful catastrophic risk (small/older systems)
ASL-2Early signs of dangerous capability; standard safeguards
ASL-3Misuse risk goes up (like CBRN uplift); strong security + deployment safeguards. Claude Opus 4 (2025) launched with ASL-3 protections
ASL-4+More advanced capabilities; stricter measures

29.2 Safety mechanisms

  • Constitutional AI training: alignment with values and principles.
  • Safety classifiers: requests are declined in some categories (cyber, bio...); in the API you see stop_reason: "refusal" and stop_details.category.
  • Server-side fallbacks: on a refusal, automatically retry on another model (opt-in).
  • Red teaming and model cards / system cards with every release.
  • Interpretability research: understanding what is happening inside the model.
  • Usage Policy: what is not allowed (weapons, CSAM, fraud...).

29.3 Prompt injection and jailbreaks

ThreatWhat it isDefense
Direct jailbreakThe user manipulates the model into breaking the rulesHarmlessness screening (pre-check with Haiku), a clear system prompt, monitoring
Prompt injectionInstructions hidden in a web page / email / tool resultTreat tool outputs as data, least privilege, human approval for risky actions
Data exfiltrationGetting the agent to send secrets outNetwork allowlist, secrets not in the sandbox (vaults), output filtering
Excessive agencyThe agent has too many permissionsPermission modes, deny rules, sandboxing

29.4 Where Claude is deployed

PlatformDetail
Claude API (Anthropic)New features first; platform.claude.com
Amazon BedrockThrough your AWS account; model IDs with the anthropic. prefix; IAM auth
Google Cloud Vertex AIGCP project + region; ADC auth; some features differ
Microsoft FoundryClaude in the Azure ecosystem; Microsoft Marketplace billing
Claude Platform on AWSAnthropic-operated, through AWS, API parity
from anthropic import AnthropicVertex, AnthropicFoundry

vertex = AnthropicVertex(project_id="my-gcp-project", region="global")
r = vertex.messages.create(model="claude-opus-5-5", max_tokens=1000,
                           messages=[{"role": "user", "content": "Hi"}])
# Bedrock: AnthropicBedrockMantle(aws_region="us-east-1"), model="anthropic.claude-opus-5-5"

29.5 Enterprise / Admin features

  • Workspaces: split API keys, limits and spend across teams.
  • Admin API: members, API keys, workspaces, usage/cost reports programmatically.
  • SSO, SCIM, audit logs, role-based access (Enterprise).
  • Data residency / inference_geo: which region inference runs in.
  • Zero Data Retention (ZDR): for eligible customers.
  • OpenTelemetry: monitoring of API calls and Claude Code usage.