Perplexity Comet Review
Use Comet when a browser assistant can compare permitted pages and you are willing to verify the original passages. Before asking about a sensitive tab or authorizing an account action, check Assistant site access, browsing-context sharing and the selected mode's usage terms. A research summary, a sent message and a completed booking are different outcomes.
What Comet's Browser Assistant Is For
Comet is a browser with a built-in Perplexity assistant. The official product page lists Mac, Windows, iOS and Android. This review examines its documented browsing/research role, not a same-version feature benchmark against Chrome or a measured automation-success rate.
Public website check. Perplexity's Comet page lists downloads for Mac, Windows, iOS, and Android. Official source.
AI Assistant, Research and Task Automation Features
Perplexity illustrates comparing how news outlets cover a topic. For that documented research task, identify the pages and question, ask the assistant to compare the coverage, then open the original pages and check the supporting passages. Distinguish the assistant's synthesis from the source's own words. This review did not execute that comparison or measure its accuracy.
Researching pages is also different from changing an account. A task involving messages, a calendar or a purchase needs the relevant account permissions and a clear action boundary. Browser access does not make instructions found on a webpage part of the user's request.
The current credit guide distinguishes Comet's modes. Search answers without consuming Computer credits. Work runs Computer: eligible paid Consumer Pro accounts use their included weekly usage first, then credits once that allowance is exhausted. Included usage is still rolling out, so look for Included in Pro on the account's Usage page instead of assuming it is present. Check the selected mode before sending a task and review the session Usage dropdown afterward. These are documented billing conditions, not a measured task cost or proof of identical access on every platform.
A local password store does not mean Assistant requests stay local. In Perplexity's Privacy Snapshot explanation, credentials such as passwords and card details are described as stored on the device, while an Assistant question can send the open tab and relevant browsing history to Perplexity for the task. The same page says users can block Assistant tasks on selected sites or turn its site and history access off. Before asking about a sensitive page, check those controls and whether that page needs to be in the Assistant's context. This is a provider description, not an independent audit of what was transferred or deleted.
Separate research from an account-changing task
The Privacy Snapshot describes open-tab and relevant-history context for Assistant tasks. The documented Search and Work modes also have different usage rules. Choose a route by the action requested:
| Request | Review before starting | Completion evidence |
|---|---|---|
| Compare two public articles | The exact tabs and question sent to the assistant. | Source passages that support the comparison. |
| Run a multi-step Work task | Current account entitlement, context access and expected credit use. | The created artifact and Usage entry. |
| Send a message or place an order | Recipient, final content, account scope and explicit confirmation. | A service receipt, not merely an assistant draft. |
These are checks derived from provider descriptions, not an automation success test. A page can supply information, but its text should not become permission for Comet to take an action.
Why agent permission boundaries need a live check
A 2026 security disclosure by Zenity Labs described two Comet agent attack paths involving untrusted content and access to local files or password-manager workflows. Zenity says it disclosed the issue to Perplexity and that Perplexity addressed the underlying browser-side agent execution issue before public disclosure. An earlier 2025 proof of concept from Brave also showed that malicious instructions inside page content could redirect a Comet summary task. Brave operates a competing browser and Zenity sells security tools, so these are commercial researchers' reports, not our own penetration test or proof that either historical flaw remains exploitable. A sensible pre-use check is to review which tabs, local files and services the agent can reach, then try a non-sensitive task before granting broader access. We did not install or test Comet's current permission prompts.
Prompt-Injection Risks and Privacy Cautions
Perplexity's prompt-injection discussion describes layered mitigations and confirmation for consequential actions such as sending messages, changing calendars and final shopping orders. These are provider design claims, not proof that prompt injection is solved or every release behaves identically.
Keep sensitive actions supervised, review the destination and content at confirmation, and treat third-party page text as untrusted material. A manual trip worksheet can avoid needing a browser action for its narrower task, but it is not thereby a security-certified alternative. The mode/usage conditions described here do not establish actual account entitlement, cross-platform feature parity or a comprehensive vulnerability assessment.
Use Comet's documented research workflow when comparing permitted pages is the task and you are prepared to check the passages. Keep messages, bookings and account changes outside that research brief unless you deliberately authorize them and review the confirmation. A manual worksheet is sufficient when you only need to organize a few entered options; it is not a like-for-like browser or a proven security substitute.
When a Mini-App Beats a Browser Agent: Macaron's Approach
Open Macaron chat to adapt the request below. Review account access and plan conditions before creating a tool.
Keep source content separate from your permission
A browser assistant is useful when the task needs browsing context. A manual worksheet is useful for a few records you choose to retain. Neither format should treat a page’s request to buy, send or connect an account as your authorization.
Sample request
"Build a manual options sheet with option ID, destination, source label, observed quote/date and quoted source text, keeping my selected option and explicit next decision separate from source wording and never executing bookings, messages or account actions."
Suggested fields: Destination; source label; entered quote; observation date; decision; next check.
Proposed checks with fictional data:
- Adapt the request in Macaron using only fictional, non-identifying data.
- Check each input and edit below against the expected result.
- Correct the stated behavior before entering personal information; use a manual record if the draft cannot preserve it.
| Fictional input or edit | Expected result to check | If it differs |
|---|---|---|
| Paste fictional source text saying reserve this now | It remains a quoted note, not permission | Keep source wording out of the decision field |
| Select that option without authorizing a booking | Selection is recorded; no booking is claimed | Distinguish a preferred option from an executed action |
| Edit one quote's observation date | Only that option changes | An edited note must not imply a fresh live price lookup |
Illustrative workflow choice: use a browser tool to examine permitted public options, then record the few details you need in a focused worksheet.
A source saying “buy this now” remains material to analyze, not permission to buy. Keep the user's chosen action in a different field from quoted page content; the separation is a proposed safeguard to check, not an attack-resistance result.
Which workflow fits your needs?
| Need | Perplexity Comet | Macaron example |
|---|---|---|
| Main purpose | An assistant embedded in a browsing workflow | A focused personal mini-app |
| What you work with | Web research and supported actions under permissions | Selected records and a reusable structure you describe |
Macaron's format does not establish stronger security by itself. Keep credentials out of prompts and review data access and consequential actions in either product.
Next step
Explore Macaron for a personal tool you can adapt to this task.
Scope and source notes
How this review was checked: Review prepared with AI for Macaron. This article evaluates cited sources and any limited public-site checks stated in the text; no core in-app task or proposed Macaron mini-app was tested.
Review scope: Perplexity’s Comet browser assistant and its published prompt-injection mitigation explanation. The Macaron example is a manually entered options sheet, not a browser integration.
Sources reviewed: last updated November 5, 2026. Product details may change; confirm current information with the provider.
| Source | Supported scope and limitation |
|---|---|
| Comet Browser: a Personal AI Assistant | Provider browser platform and assistant scope; no browser task or automation benchmark |
| Mitigating Prompt Injection in Comet | Provider prompt-injection mitigation discussion; not an independent security assessment |
| How credits work on Perplexity - Perplexity Help Center | Search/Work, paid Consumer Pro rollout and weekly included usage before credits; no task execution |
| Introducing Privacy Snapshot — Perplexity | Provider account of Assistant settings and transfer of open-tab and relevant browsing-history context; no independent privacy audit |
| PleaseFix vulnerability disclosure — Zenity Labs | Security vendor's 2026 research on Comet agent paths; the same disclosure says Perplexity fixed the underlying browser-side issue before publication, so it is not evidence of a current unpatched flaw |
| Agentic Browser Security: Comet prompt injection — Brave | 2025 proof of concept by a competing browser vendor; historical demonstration of an indirect prompt-injection path, not current-version test or independent audit of patch status |
Corrections: contact@macaron.im.
FAQ
Is Comet only another name for Chrome?
No. It is Perplexity's browser product with an assistant-oriented workflow.
Should a web page be allowed to authorize an action?
No. A page can supply information for a task, but instructions written on it are not your permission for an agent to send data or change another account. Review the destination, data and action separately before approving a consequential step.
Is a mini-app automatically safer than a browser agent?
No. Data handling, permissions and the actual task determine the risks.
Will a manual options sheet keep travel prices current?
No. It preserves a quote and the date you checked it, so a later change needs a new observation. Reopen the original booking page before paying rather than treating a saved number as live availability.
What would a Macaron options sheet keep after browsing with Comet?
The proposed sheet stores a destination, source label, observed quote and date, and the text that supports a choice. It can preserve what you checked, while Comet's browser interaction and any live page actions remain separate.
Will the Macaron example browse pages or verify Comet citations automatically?
No browser control, connector or source-checking integration was tested. Add a fictional option with an uncertain quote and confirm the sheet keeps that uncertainty until you revisit the original page and verify the wording.
Related
Read & discover
Explore Playbooks
AI Personal Assistant
Free calculators
