Skip to main content

The EU AI Act gets teeth. What browser-agent teams need

The AI Act's high-risk dates moved. Its transparency rules did not. Browser-agent teams still need evidence.

Estimated reading time: 10 min read
Updated
A red route crossing a raised gate toward a dark monolith marked with the EU stars

Most AI Act rules have applied since August 2, 2026. The high-risk deadlines moved later. Annex III high-risk rules apply from December 2, 2027. Annex I product rules follow on August 2, 2028. Article 50 transparency rules already apply. Your duties depend on the use case, not the browser.

The reading below is engineering guidance, not legal advice. It maps each date to evidence your team can produce.

What applies now, and what moved

The AI Act entered into force on August 1, 2024. Its obligations arrive in phases. The AI Omnibus is Regulation (EU) 2026/1744. It entered into force on July 27, 2026. It moved the high-risk dates.

Dates below follow the European Commission's AI Act page. They also match the AI Act Service Desk timeline.

DateWhat appliesWhy browser-agent teams care
August 1, 2024The AI Act entered into force.The phased timeline starts.
February 2, 2025Prohibited practices and AI literacy duties apply.Check agent tasks against the banned practices.
August 2, 2025Obligations for general-purpose AI models apply.These duties target model providers.
August 2, 2026Most remaining rules apply, including Article 50. Enforcement of applicable rules starts.Transparency duties can reach agents that deal with people.
December 2, 2026New prohibitions apply. The Article 50(2) grace period ends. It covered certain synthetic-content systems already on the market.Relevant if your agent generates synthetic content.
December 2, 2027Annex III high-risk rules apply.Relevant for agents in a listed high-risk use.
August 2, 2028Annex I high-risk rules for regulated products apply.Relevant only for AI built into regulated products.

General-purpose AI obligations applied from August 2, 2025. Those duties target the providers of the models. Many browser-agent teams call those models through an API. They add prompts, tools, credentials, browser access, and business logic.

Enforcement of applicable rules started on August 2, 2026. The AI Office and national authorities now share that work.

What are the EU AI Act penalties?

Fines scale with the violation, under Article 99. Prohibited practices can draw up to EUR 35 million. The alternative cap is 7 percent of worldwide annual turnover. The higher figure applies. Listed operator duties reach EUR 15 million or 3 percent. Article 50 transparency duties sit in that tier. For SMEs, the lower figure applies. Each figure is a ceiling, not a running cost.

The compliance slide is not the control. Evidence is the control. Show where the agent ran and what it touched. Show what it retained and how the session ended.

For engineering leaders, the timeline changes priorities more than architecture. You do not need a regulation-themed rebuild. You need a map of the system you already run. Then trim the data and persistence you do not need.

Does the Act reach your browser agent?

Start by separating the model from the deployed system. A general-purpose model may plan, summarize, or pick actions. Your application decides everything around it. It sets the tools, the pages, and the credentials. It also sets approvals and output destinations.

Model and application are different control surfaces. The provider owns the model layer. The deploying team owns much of the surrounding workflow. Obligations vary by role and by system category. A vendor's GPAI work does not classify your deployment.

Do not classify from the product label alone. “Browser agent” describes an interface, not a use case. Compare an agent that summarizes public documentation. Then picture one inside hiring, lending, healthcare, or public services. The data and decisions differ sharply. The Commission's high-risk examples include CV sorting. Credit scoring is another listed example. Workflow context is the useful starting point for counsel.

Deploying teams still need answers on transparency and risk. Who knows an agent is acting? Which actions require approval? What happens when the model makes a bad choice? Which records let someone reconstruct the run? Exact legal duties depend on the deployment. Answer the engineering questions before that determination.

Do Article 50 transparency rules apply to AI agents?

Article 50 can reach some agents. Its rules have applied since August 2, 2026. Article 50(1) covers systems that interact directly with people. Providers must make that interaction known, unless it is obvious. Article 50(2) requires machine-readable marks on synthetic outputs. Deployers must also disclose deepfakes. An agent that messages customers may trigger the first duty.

Stack layerOperator questionUseful evidence
GPAI modelWhich provider and model version handled the task?Provider record, model identifier, and change history
Agent orchestrationWho set the objective, tools, and approval gates?Agent configuration, tool policy, and named owner
Browser executionWhere did the page run, and what was isolated?Session record, execution location, and boundary description
IdentityWhose credentials entered the session?Credential source, granted scope, and approval record
Outputs and logsWhat left the session or survived its close?Destination log, retention setting, and teardown event

The stack table is not a legal classification test. It is a responsibility map. Without it, each vendor may own one narrow layer. Then nobody owns the assembled browser-agent system.

The assembled system is what your users experience. It is also where data moves between boundaries. A prompt enters the model. A tool opens a browser. A credential authenticates a page. A screenshot enters a log. Each handoff needs an owner and a retention answer.

The engineering posture: least data, least persistence

Picture a browser agent working a supplier portal. It signs in, reads invoices, and writes a summary. The run may touch names, records, cookies, and tokens. Then a reviewer asks what remains after the task ends.

The model provider cannot answer that question. The answer lives in your own stack. Check orchestration, browser infrastructure, identity, logging, and retention. Pieces that grew independently turn review into a hunt.

A browser session accumulates more than page content. It can hold cookies, local storage, form data, and downloads. Tokens, screenshots, and rendered page text can stay too. A long-lived profile carries those artifacts forward. A shared agent host can expose them to later work.

Every retained artifact creates another question. Why was it collected? Who can read it? How long does it stay? Which process deletes it? Broad capture and indefinite retention multiply those questions. They also multiply failure modes.

The cleaner default is least data and least persistence. Give the agent only the credential its task requires. Keep durable logs only for a defined purpose. Operations, security, and governance each qualify. End browser state when the job ends.

A separate isolated browser can support that posture. It moves the browser step away from a daily profile. Cookies, files, downloads, logs, and persistence still need checking. They depend on the product contract and configuration.

Legba's agent skill gives agents routing and isolated browser sessions. Shield opens a page in an isolated browser off your device. Verify permissions, transfers, logs, retention, and cleanup first. Then rely on that boundary.

No infrastructure component can make a deployment compliant. Compliance depends on the whole system and its use. Data, operators, and applicable duties matter too. Browser isolation is one risk-reduction choice. Its minimization value depends on measured retention and transfers.

Deletion also needs proof. A “close” button proves nothing on its own. Confirm the container ended and the volume disappeared. Confirm revoked tokens too. Capture the platform event that marks teardown. Test timeouts and interrupted runs as well. Make the evidence available without exposing the agent's reasoning.

Questions to ask your stack this quarter

Turn the architecture review into a short operator interview. Include engineering, security, identity, privacy, and procurement. Bring the business owner too. Record any gap between team answers before choosing a fix.

  • Where do agent sessions execute? Name the service and execution location. Describe the isolation boundary and what it can reach.
  • What persists after a run? Check cookies, local storage, cache, files, and screenshots. Then check prompts, DOM captures, tool outputs, and backups.
  • Whose credentials ride along? Separate user credentials from service identities. Track delegated tokens and secrets the orchestration layer injects.
  • What gets logged? List URLs, prompts, outputs, tool calls, and screenshots. Add network events, approvals, and errors. Name who can access each record.
  • Can you show session teardown? Produce a machine record for close, timeout, crash, and cancellation. Verify that the underlying browser state is gone.
  • Who owns agent changes? Record who can switch models or edit prompts. Include who adds tools, expands destinations, or changes retention.
  • What requires a human decision? Require approval before actions that send data or spend money. Do the same for record changes and access grants. OWASP's excessive agency guidance recommends this too.
  • What happens after failure? Define how the run stops and credentials get revoked. Name the records that remain and who reviews the event.

Do not accept “the vendor handles it” as complete. Ask which vendor, which layer, and which contract term. Then find the handoff back into your own stack. Responsibility often shifts at the API response or browser launch. Credential injection and log export are common handoffs too.

Run the checklist against a real workflow. Trace a recent agent task from request to teardown. Architecture diagrams show intended paths. Session records show what actually happened. Keep both, and reconcile them when they disagree.

Practical next steps

  1. Inventory every agent. Unsanctioned tools often escape normal review. The shadow AI field guide explains why. Record each agent's owner, model, tools, data, and credentials. Add destinations, retention, and the teardown mechanism.
  2. Inventory workflows, not just vendors. One model can serve public research and sensitive work. Those deployments deserve different access, logging, and review. A vendor list alone hides that distinction.
  3. Isolate browser execution. Match isolation to the authority each task receives. Legba's agent skill is one option for the browser step. Compare agent plans before you commit.
  4. Separate identity from convenience. Do not let an agent inherit a daily profile. Easier login is not a reason. Use the smallest credential scope the workflow supports. Put approval around credential use and state changes. Revoke access when the declared job ends.
  5. Set retention deliberately. Keep the evidence security, operations, and governance need. Remove browser residue that serves no durable purpose. Document why each surviving artifact exists and who owns it. Note what ends its retention.
  6. Build an evidence pack per workflow. Include the owner, system map, and agent configuration. Add the credential path, approval points, and log destinations. Add retention settings and teardown proof. Update the pack when models, tools, or boundaries change.

The later high-risk dates are not a pause. Transparency duties and enforcement already apply. Counsel decides which obligations reach your deployment. Engineers make the system legible enough for that call.

Make every answer concrete. Say where the session runs and which credential it carries. Say where data goes, what remains, and how it ends. Less data means less exposure. Less persistence means less residue. Clear teardown means better evidence.

How this reading was made

Dates come from the European Commission's AI Act pages. Those pages reflect the AI Omnibus amendments. Article summaries follow the Commission's AI Act Service Desk. Sources were checked on October 30, 2026.

The reading covers browser-agent engineering, not every duty. It does not classify any specific system. Confirm your obligations with counsel. Legba sells the agent skill and the extension mentioned here.

Related reading on agent inventory, isolated execution, and disposable browser infrastructure.

Free for 30 days. No card required.

Try it on the next page you do not trust

Legba is a Chrome extension with two modes. Ghost gives you a private browser route. Shield opens a page in an isolated browser off your device.

Visit the security and trust center

About the authors.

Try it on the next page you do not trust

Legba is a Chrome extension with two modes. Ghost gives you a private browser route. Shield opens a page in an isolated browser off your device.

Free for 30 days. No card required.

Ghost. A private route for your browser. Shield. An isolated browser, off your device. Choose the mode. Close when finished.