v0.2.0 — Multi-Turn Attack Chains

What's new

Multi-Turn Attack Chains

The biggest gap in OSS LLM security testing: attacks that span multiple conversation turns. Real jailbreaks happen through gradual escalation — single-shot fuzzers miss the entire class.

12 built-in chains across 3 categories:

• Jailbreak (4 chains) — researcher rapport → persona switch, authority claim escalation, roleplay ramp-up, hypothetical frame collapse
• Data extraction (4 chains) — capability probe → system prompt extraction, incremental segmentation, false memory injection, echo pattern escalation
• Injection (4 chains) — gradual instruction erosion, context poisoning, delimiter injection, goal drift

Each chain has conditional branching: if the model refuses one approach, it tries a different angle automatically.

Wizard upgrade

After importing a curl command, you now choose your attack mode:

• Single-shot only — 165 targeted attacks
• Multi-turn only — 12 chain attacks
• Both — full coverage

CLI flag

```bash
promptfuzz test https://your-api.com/chat --multi-turn
promptfuzz scan --target https://your-api.com/chat --multi-turn --output report.html
```

Install

```bash
pip install promptfuzz==0.2.0
```