Skip to content
FlashTowerFlashTower home

AI Security · Perplexity

Perplexity security for business research

Perplexity is an AI answer engine that combines language models with live web search. It is popular for research because it cites sources. The security questions concern what employees include in queries and uploads, and what the product is allowed to do when it browses for them.

Last reviewed: September 2026

What Perplexity is

Perplexity answers questions by searching the web and summarizing results with citations. It offers free and paid individual plans and business offerings with administrative controls. The company has also introduced agentic browsing capabilities that can navigate sites and take actions for a user.

Verify current terms

Vendor plans, features, and contractual terms change often. Treat the descriptions on this page as a starting point and confirm the current terms, admin settings, and data-handling commitments for your specific plan directly with the vendor before relying on them.

What Perplexity can access

  • Queries — which frequently reveal strategy: acquisition targets, litigation research, product plans, or client names.
  • Uploaded files for analysis alongside web results.
  • Connected sources, where business plans support searching internal repositories alongside the web.
  • Browsing sessions, when agentic features act within a user's browser, potentially including sites where the user is logged in.

Key risks

Sensitive context in queries

Research queries feel harmless, but a sequence of searches can reveal confidential intent. For law firms, investment teams, and corporate development, the query itself may be confidential information.

Unapproved individual accounts

Perplexity is often adopted individually and expensed. Without a business plan, the organization has no retention controls or visibility.

Agentic browsing and prompt injection

An AI that reads arbitrary web pages and can also act in logged-in sessions is the textbook prompt-injection scenario: a page can contain instructions that attempt to make the agent visit other sites, submit data, or reveal information from the session. Treat agentic browsers as privileged software and restrict them until you have reviewed their safeguards.

Source reliability

Citations make answers easier to verify, not automatically correct. Decisions should rely on the cited primary sources, not the summary.

Recommended controls

  1. Decide whether Perplexity is approved; if so, standardize on an administered business plan with SSO.
  2. Document the plan's retention and training terms in your vendor register.
  3. State in policy that client names, deal names, and confidential matter details must not be used in queries on unapproved tools.
  4. Restrict agentic browsing features, especially on devices with access to sensitive web applications.
  5. Detect individual subscriptions through expense review.

How FlashTower assesses this

When you add Perplexity to a FlashTower Scan, the Risk Engine evaluates how it is actually used in your organization — account type, data accessed, capabilities, autonomy, credentials, logging, and vendor terms — against a versioned, deterministic ruleset. These rules are the most relevant for Perplexity. Each fires only when its stated condition is met. See how scoring works.

Identity

  • FT-IDN-001High

    No controls on unapproved (shadow) AI tools

    Unapproved AI tools are neither blocked nor monitored, or this cannot be confirmed.

  • FT-AUT-010Medium

    Team-wide use of consumer AI accounts

    More than ten people use this AI system on a free or personal consumer plan.

Sensitive Data

  • FT-DAT-011Critical

    Highly sensitive data in a consumer AI account

    A free or personal consumer AI plan is used with health data, payment card data, privileged legal material, or credentials.

  • FT-DAT-012High

    Sensitive business data in a consumer AI account

    A free or personal consumer AI plan is used with personal information, financial data, customer records, or source code.

External Communications

  • FT-COM-001High

    AI sends or publishes externally without review

    The AI system can send email/messages or publish content without per-action human approval.

Third-Party Integrations

  • FT-3PI-001High

    Prompt-injection exposure: web access combined with sensitive data or actions

    The AI system can browse the web or call external URLs and also has sensitive data or high-impact capabilities.

Governance

  • FT-GOV-010High

    Unapproved AI tool handles sensitive data

    The AI system is not approved (employee-adopted) and accesses sensitive data.

  • FT-GOV-011Medium

    Unapproved AI tool in use

    The AI system is not approved and does not access sensitive data.

Vendor Risk

  • FT-VEN-010Medium

    Vendor security review not completed for sensitive-data AI system

    A vendor-hosted AI system accesses sensitive data and no vendor security review is recorded.

Frequently asked questions

Is Perplexity safe for business use?

It can be used for general research with an approved, administered account and clear rules about confidential information in queries and uploads. Individual accounts with sensitive queries and unrestricted agentic browsing are the main concerns.

Are search queries sensitive data?

Often, yes. Queries about specific companies, people, or legal matters can reveal confidential intent. Treat them with the same care as documents.

What is the risk of AI browsers?

AI agents that browse the web read untrusted content that can include hidden instructions. If the same agent can act in logged-in sessions, those instructions may attempt to misuse that access. Restrict and review these features before approving them.