Perplexity AI is a San Francisco-based company founded in 2022, providing an AI-powered answer engine that combines LLMs with real-time web search for cited responses. Enterprise platform offers internal knowledge search, integrations, and security features like SOC 2 Type II, GDPR/HIPAA compliance.
Perplexity AI is a San Francisco-based company founded in 2022, providing an AI-powered answer engine that combines LLMs with real-time web search for cited responses. Enterprise platform offers internal knowledge search, integrations, and security features like SOC 2 Type II, GDPR/HIPAA compliance.
Perplexity Computer: autonomous digital worker coordinating multiple models/sub-agents for multi-step tasks (research, code, analyze, build sites); Comet Browser for workflows; tool integrations, browser actions, API access, file/system management in isolated environments.
Copyright lawsuits: NYT/Chicago Tribune (Dec 2025), Reddit scraping (Oct 2025), Dow Jones/NY Post (2024). Stealth crawling evading blocks (Cloudflare Aug 2025). Hallucinations in Deep Research (Feb 2025), fake news in responses (2024). Prompt leak (Dec 2025). Phishing on Comet launch (Oct 2025). Service outages reported.
Detailed breakdown of every risk category for enterprises deploying Perplexity models in agentic AI workflows.
Despite citations, instances of fabricated sources/dates (Deep Research hallucinated 2025 reports), fake article content (Dow Jones suit), citing flawed data like racist IQ studies. Relies on underlying LLMs; reduced but not eliminated by search grounding.
Consumer: collects queries/IP/device for training. Enterprise: no training on customer data, configurable retention, SSO/SCIM. Potential leakage in API/prompts if sensitive data input; compliant but user-dependent.
Surfaced debunked 'race science' IQ data from Lynn in national IQ queries (Wired Oct 2024), reflecting upstream LLM/web biases despite mitigations.
Prompt injection leaked system prompt (2025). Jailbreaks on hosted models like Claude (2025). Stealth scraping issues. No major breaches, but agent tools/browsers increase attack surface.
Agentic workflows (Computer/Comet) with browser/tools/filesystem access in isolated envs; risks if misconfigured or compromised (e.g. malicious prompts executing unintended actions). No specific incidents.
Orchestrating 20+ models risks API changes, latency, safety mismatches. User reports of agent failures in complex tasks (final 20% issues). No quantified drift data.
GDPR compliant with data rights process. Ongoing copyright suits may set precedents. EU AI Act: high-risk agentic use may require conformity assessments. HIPAA for eligible customers.
Multiple suits alleging unauthorized scraping/training on publisher content (NYT, Reddit, News Corp). Revenue-sharing proposed but disputed.
Supports video gen (Veo 3.1, 15/mo Enterprise Max); general LLM misuse potential, but text/search-focused. No specific incidents.
Cyber liability for data breaches/leaks; E&O for erroneous agent outputs/hallucinations causing harm; IP infringement coverage; AI-specific riders for model misuse, regulatory fines. High due to agentic tools.
Barmenia Gothaer, Berkshire Hathaway Automotive, Volvo Cars US, Galeries Lafayette, En; 1,300+ tracked.