Security and Privacy at the SDK Boundary

What leaves your process

Every SDK call sends prompts, system instructions, attached files, and tool definitions to Anthropic's servers (or Bedrock/Vertex). The reverse-direction surface is what the model can do — read files, run shell, hit URLs. Both directions are part of your security model.

Secrets discipline

API keys belong in environment variables and secret managers, never in code, never in prompts, never in tool descriptions. JSONL logs are particularly dangerous — value-pattern keys (sk-ant-...) leak into git history if you log the system prompt or request body unredacted.

Data residency and retention

Anthropic's default API does not train on your inputs. Bedrock and Vertex respect their cloud's data residency commitments. Zero Data Retention (ZDR) is a feature available on certain plans and changes the privacy story for regulated workloads. Know what your account has signed up for before classifying data as 'safe to send'.

Principle: Treat the SDK boundary as a public boundary. What you send and what you log are both audit trails.

Code

Redact API keys from any structured log·python

import re

KEY_PATTERN = re.compile(r"\bsk-ant-[A-Za-z0-9_-]{20,}\b")

def redact(text: str) -> str:
    return KEY_PATTERN.sub("sk-ant-***REDACTED***", text)

# Always run logged payloads through redact() — even in private repos.
log_line = '{"prompt": "hi", "key": "sk-ant-api03-abcdef0123456789"}'
print(redact(log_line))

Pre-commit hook to refuse staged files containing key patterns·bash

#!/usr/bin/env bash
# .git/hooks/pre-commit
set -euo pipefail
if git diff --cached -U0 | grep -E 'sk-ant-[A-Za-z0-9_-]{20,}' >/dev/null; then
  echo 'refusing to commit: detected an Anthropic key pattern in staged content'
  exit 1
fi

Security and Privacy at the SDK Boundary

What leaves your process

Secrets discipline

Data residency and retention

Code

External links

Exercise

Progress

Comments 0