AI Agents for Lawyers

Glossary

Computer Use / Browser Agent — Definition for Lawyers

Computer use lets an AI agent see a screen and click, type and browse for you. Definition, Claude for Chrome, ChatGPT Work, Perplexity Computer and the legal risks.

By Dr. Niklas SchmidtUpdated 13 November 20264 min read
On this page
  1. The loop, and the three products that run it
  2. What lawyers point it at, and whose login it is
  3. The risk that does not go away
  4. Controls before you switch it on

Computer use is an AI agent's ability to operate software the way a person does: it looks at the screen or the page, decides where to click, types, scrolls and reads the result, then repeats. A browser agent is the same capability confined to a web browser. For lawyers it means an agent can work in portals and registers that have no API, under your own login.

The loop, and the three products that run it

Every computer-use agent runs one cycle: perceive (a screenshot or the page's structure), decide (which element to act on), act (click, type, scroll), perceive again. Because it reads whatever is on the page, it can work in a decade-old land-registry portal as readily as a modern app. Because it reads whatever is on the page, the page can also tell it what to do, which is the security problem below.

ProductAvailable sinceWhere it runsWho gets itBuilt-in controls
Claude for ChromeGA 18 December 2025Your own browser, your logged-in sessionsPro, Team, EnterpriseAdmin allowlist and blocklist of sites
ChatGPT Work9 July 2026OpenAI's cloud, including a cloud browserPaid ChatGPT plans, meteredPlan mode, approval gates, Compliance API audit trail
Perplexity Computer25 February 2026Isolated cloud environment with a real browser and file systemMax subscribers onlyEach task in its own sandbox

Sources: Claude for Chrome, ChatGPT Work, Perplexity Computer. GPT-6 Astra, OpenAI's computer-use model, is switched off by default in ChatGPT Enterprise until an admin enables it (9to5Mac, 4 September 2026): the general counsel, not the associate, decides.

The column that matters for confidentiality is "where it runs". A local agent uses your sessions and your machine. A cloud agent means a vendor's computer holds the credentials and every page it visits, which for European firms raises Article 28 GDPR and professional-secrecy questions before any productivity question. The desktop cousin is covered in Claude Cowork for lawyers.

What lawyers point it at, and whose login it is

The honest use cases are clerical and portal-bound: pulling current extracts for thirty group companies from a company register and tabulating the directors; re-keying matter data into a practice-management system with no import function; pre-filling an e-filing form from a draft. The agent saves an afternoon of clicking; a person checks the result.

Two legal points sit on top. Many registers and portals prohibit automated access in their terms. And on 4 August 2026 the Ninth Circuit vacated Amazon's injunction against Perplexity's Comet browser, reasoning that "It is the user who 'accesses' Amazon's computers, with the help of the Assistant", at least where the agent runs on the user's device (Amazon v Perplexity, summarised). That cuts both ways: the agent is not a trespasser, and any breach of a portal's terms lands on the lawyer whose login it used.

The risk that does not go away

A browser agent cannot reliably tell content it was asked to read from instructions hidden inside it. Zenity's "PleaseFix" research (disclosed 3 March 2026, presented at Black Hat on 5 August 2026) showed zero-click hijacks of Claude in Chrome, Gemini in Chrome, Comet, Atlas and Copilot Edge through emails, calendar invites and web pages; in one demonstration "summarise my emails" ended with Gmail and Drive contents exfiltrated and Slack, X and Claude passwords reset (Dark Reading). Zenity's verdict: "The fix isn't to keep asking the AI to behave. It's to put hard limits around the agent that the agent cannot override."

Anthropic's figures say the same from the other side. Mitigations cut attack success in Claude for Chrome from 23.6% to 11.2% across 123 scenarios (December 2025); its November 2025 research reported about 1% against an adaptive attacker with Opus 4.5, while stating that "No browser agent is immune to prompt injection" (Anthropic). One poisoned page in a hundred is fine for a shopping assistant and poor for an agent logged into your DMS. The mechanics are in prompt injection; the desktop version of the lesson is the February 2026 Cowork incident, where an approved "delete temporary files" became an rm -rf over 15 years of family photos (Futurism).

Controls before you switch it on

  1. Separate browser profile. Nothing logged in except the one site the task needs. No DMS, no email, no client portals.
  2. Allowlist, not blocklist. Only an allowlist survives a page you did not anticipate.
  3. No inbox, no calendar. Mail and invites are the attack channel in every published hijack.
  4. Read and pre-fill only. A person presses Submit, Pay, File and Send. Write that into the firm AI policy.
  5. Synthetic matter first. Watch the first ten runs.
  6. Know what is logged. Anthropic's Compliance API returns Claude-in-Chrome session transcripts for Enterprise customers from 18 September 2026, with six-year default retention, and "capture must be enabled before local activity occurs" (Anthropic docs). An audit trail and a discoverable record at once.

Treat a computer-use agent as a temp with your password. The wider threat model is in agent security for lawyers, and Module 4 of the AI Lab for Lawyers course covers browser and computer use alongside n8n, Make and Zapier, using browser-based tools only.

Frequently asked questions

What is computer use in AI?

Computer use is a capability that lets a language model operate a computer the way a person does. The model receives a screenshot or the page structure, decides on an action such as click, type or scroll, performs it and looks again. Repeated in a loop, that lets an agent complete tasks in any application, including websites and portals that have no programming interface.

Is Claude for Chrome safe for law firms?

It is usable with controls, not safe by default. Claude for Chrome has been generally available on Pro, Team and Enterprise plans since 18 December 2025 and lets admins allowlist or blocklist sites. Anthropic reports it cut prompt-injection attack success from 23.6% to 11.2% in its own tests, but also states that no browser agent is immune. Use a separate profile with no client systems logged in.

Can an AI browser agent file documents with a court?

Technically it can navigate an e-filing portal, but no firm should let it press Submit. Court rules and bar guidance put personal responsibility for every filing on the lawyer, and Anthropic's own legal plugins hold a gate before anything is sent or filed. Let the agent prepare and pre-fill; a named lawyer reviews and files.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion