GitHub shipped a Secret Scanning update on October 13. One line:

Two Anthropic key types → blocked by push protection by default
+ new Salesforce and Cloudsmith secret types detected

Teams using AI coding tools should not skip this. Agents now commit config files, which changes how keys leak.

Two different things: detection vs blocking

Detection (Secret Scanning)
  → finds keys already in history, raises alerts
  → works where Secret Scanning is enabled

Blocking (Push Protection)
  → stops the push before the key lands in history
  → Anthropic's two types now included by default

The new Anthropic types cover service and user keys. Commits containing them are blocked at push time in repos with Secret Scanning on. Salesforce and Cloudsmith types join detection coverage too.

Context helps: a week earlier, on October 7, GitHub released a purpose-built leak-detection model that reads surrounding code instead of matching token shapes. Detection widens with AI, blocking tightens by default. A two-sided strategy.

Why Anthropic keys, why now

Because AI coding agents write code and touch configs along the way:

Old leak: a human accidentally adds .env
New leak: an agent-generated config carries a key
          through commit and push without review

Anthropic keys are among the most-handled credentials now that Claude Code-style tools are daily drivers. Once a key lands in history, rotation costs follow, plus secondary abuse through resellers and proxies. Blocking at push is the cheapest response.

CodeBridge mini lab: a 10-minute repo check

[ ] Secret Scanning on for the org and repos?
[ ] Push Protection on (repo, org, or enterprise — know which)?
[ ] Bypass reasons required + bypass alerts visible in Security tab?
[ ] .env and key files in .gitignore (first line, separate from blocking)?
[ ] Rotation owner and procedure defined for a leak?

The commonly missed item is coverage: protection only helps where Secret Scanning is enabled. Check coverage before trusting the default. Free public repos block automatically; private repos depend on plan.

Pair it with a local habit:

Skim git diff --cached before pushing
→ any sk-ant-… strings visible?
→ hardcoded values in newly created configs?
→ agent-created files get one human pass

Layer it with existing gates

This change shines combined, not alone:

push protection (block at push)
+ /security-review (Copilot pre-push review)
+ context-aware detection (finds what slipped into history)
→ layered across commit → push → PR

Keep the practice from our context-aware detection guide: track false positives separately from real saves, or developer trust erosion stays invisible.

Conclusion: from incident response to push gates

Key leaks are a push-gate problem, not an incident-response hobby.

When agents commit, platform blocking beats human attention. A 10-minute settings check plus reviewing agent-created files lowers the next leak's odds today.

Further reading

References

Go deeper with a course

To wire Agent and Plan modes into real projects with security gates attached, learn to use Copilot beyond autocomplete.