Claude Bug Bounty Payloads & Test Cases (2026)

Program scope (HackerOne)

• Universal jailbreaks against the latest Claude (Sonnet 4.5 / Opus 4) — up to $25k for novel CBRN bypass.
• Claude Code: command injection, sandbox escape, credential exfil — up to $20k.
• Computer Use / Browser Use: cross-site action via screenshot injection — up to $15k.
• MCP reference servers: tool poisoning, transport vulns — up to $10k.

Universal jailbreak test cases

Anthropic specifically wants jailbreaks that work across the eight CBRN-adjacent categories with a single template. Persona laundering + AFL + MSJ chain is the current frontier.

• Test all 8 categories from the Acceptable Use Policy in one report.
• Include 5+ trials per category showing >80% success.
• Provide a working harness (Python) the triage team can re-run.

Claude Code test cases

• README.md poisoning -> arbitrary shell exec on first /init.
• MCP transport: stdio framing confusion -> command injection.
• Cached credentials in ~/.claude leaking via tool error messages.
• .claudeignore bypass via symlink chains.

Computer Use test cases

• Screenshot injection that triggers OAuth consent on attacker.com.
• Drag-and-drop credential exfil via clipboard.
• Cross-tab session hijack when user authenticates mid-session.

Report write-up template that gets triaged fast

1. One-sentence impact in the title.
2. Reproducible harness (Python + minimal deps).
3. 5+ trial transcripts (sanitized).
4. Mapping to the Acceptable Use Policy section.
5. Suggested mitigation (optional but accelerates triage).

FAQ