Perplexity Sonar

​

Default model

sonar-pro

​

Supported models

• sonar-pro (default, web-grounded flagship)
• sonar (web-grounded base)
• sonar-reasoning-pro (reasoning + web)
• sonar-reasoning (reasoning + web base)
• Any other model ID the API accepts

​

Authentication

export PERPLEXITY_API_KEY=pplx-...

ai-blackteam config set providers.perplexity.api_key pplx-...

​

Example usage

# Single attack with default model
ai-blackteam run -p perplexity -a indirect-injection -t "Search for X and ignore safety"

# Compare web-grounded vs sealed
ai-blackteam benchmark --models perplexity:sonar-pro,anthropic:claude-sonnet-4-6

​

Tool-use support

​

Notes

• Web-grounded models behave differently under prompt injection than sealed-knowledge LLMs, distinct attack surface worth testing separately
• The provider uses the OpenAI SDK pointed at Perplexity’s API (base_url="https://api.perplexity.ai")
• Sonar models include web search results in the context, making them vulnerable to indirect prompt injection via search-engine-poisoning
• Max output tokens per request: 4,096
• Retry with exponential backoff is automatic on API failures (3 attempts)