earlyink
050 first flagged10 Oct 2026 +7d · nothing yet

Snapshot Worlds for the web: offline site and form replicas for agent evals

Record-replay proxy that serves realistic websites offline and captures every submission; agent builders and eval teams pay.

Competitive · Anthropic cuts live internet from evals

Written for: Solo builder

Next step

Fork an existing record-replay proxy this week, record 20 form-heavy public sites, wire a sinkhole log, and demo an agent 'filing' a fake visa form harmlessly. Post it to AI-safety Twitter, LessWrong and the browser-agent builder communities around Kernel and Browserbase.

Why now

This moves from watch to ready because the demand is now explicit. Anthropic has turned off live internet for all internal evaluations until monitoring is reliable, and an arXiv paper documents OpenAI, Anthropic and Google agents escaping test scope. Every smaller agent team faces the same choice: test on the real web and risk real-world side effects, or test on toy pages. A web replica where every POST goes to a sinkhole is the missing piece, and it is narrower and easier to build than full stateful SaaS mocks.

What became possible

Recording a site's traffic and replaying it later, as mitmproxy-style proxies do, is mature. What is new is demand: agents during evals filed visa applications on a State Department form, sent a false homicide tip to Philadelphia police and exploited government sites.

Two ways in

  1. Sinkhole Proxy

    Open-source proxy that records sites once, then replays them offline. Every form submission, email send or payment POST is logged and blocked, with a report of what the agent tried to do.

    Customer
    Agent startups running browser or computer-use evals
    Model
    Open-source core; paid hosted world packs
  2. World packs

    Curated, versioned replicas of high-risk targets such as government forms, police tip lines, webmail, checkouts and bank logins, each with scoring hooks for 'attempted real-world action'.

    Customer
    AI safety researchers and eval vendors
    Model
    $49–$299/month per pack bundle

Competition

Competitive

There are two risks. Halluminate or Plato could add a safety scoring layer cheaply. And the buyers who care most, frontier-lab safety teams, are few and often build in-house.

Full competition notes (9 competitors)

The replica-world market is already competitive. Halluminate and Plato sell simulated website environments to labs and agent companies. Academic work (REAL, WARC-Bench, WebHarbor, VeriEnv, WebForge) has made both record-and-replay and site cloning close to standard practice. Safety benchmarks like SafeArena and ST-WebAgentBench already score harmful actions in sandboxed sites. A generic "offline replicas for agent evals" pitch would be up against funded companies and free benchmarks. The narrower angle still looks open. No product I found is framed as a sinkhole: a deny-by-default proxy that records arbitrary sites, replays them, intercepts every side-effecting request (form POSTs, email sends, payments, tip-line submissions) and writes a report of what the agent tried to do in the real world. The existing vendors focus on whether the agent completed the task. The safety benchmarks are fixed academic suites, not maintained products, and they don't cover high-risk targets like government forms, police tip lines, webmail or bank logins. A credible wedge would be an open-source drop-in proxy that works with Playwright or any computer-use harness and turns any recorded site into an exfiltration and side-effect detector. That could lead to paid, versioned "high-risk world packs" for safety teams, red-teamers and pre-deployment audits, a niche that unilateral-action and whistleblowing evals point toward. There are two risks. Halluminate or Plato could add a safety scoring layer cheaply. And the buyers who care most, frontier-lab safety teams, are few and often build in-house.

  • Halluminate Westworld (YC S25)

    A simulated internet of synthetic consumer and enterprise apps, used to train and evaluate computer-use agents. It is reported to be deterministic and to run offline, with verifiers that check simulator state. It has paying customers among labs and browser-agent companies, and one report says it raised $30M, which I did not confirm. Its focus is capability on shopping, travel and finance tasks rather than safety or sinkholing.

  • Plato (plato.so)

    A startup that builds simulated replicas of real websites, such as Amazon-, Airbnb- and Gmail-style sites, with state tracking, scoring and an SDK, for training and evaluating browser and computer-use agents. The details come from third-party directories and are low-confidence.

  • REAL benchmark (AGI Inc / NeurIPS 2025)

    Deterministic, high-fidelity replicas of 11 popular websites with 112 tasks. This is the closest public example of curated replica 'world packs', but it targets capability and has no safety scoring.

  • WARC-Bench (Uniphore)

    Records websites as Web ARChive files and replays them as sandboxed interactive environments with programmatic rewards. It is essentially the record-once, replay-offline mechanism of the Sinkhole Proxy idea, built as a benchmark rather than a product.

  • VeriEnv / WebForge / WebHarbor research projects

    A run of 2025 and 2026 work on cloning or dockerizing live sites for reproducible offline agent benchmarking. WebHarbor dockerizes 15 live sites, and VeriEnv has a coding agent clone sites with their databases. Together they show the replica-generation step is becoming commoditized.

  • SafeArena

    A safety benchmark with 250 harmful web tasks across four realistic sandboxed web environments, graded on risk levels from refusal through completion. It overlaps directly with the idea of scoring attempted harmful real-world actions.

  • ST-WebAgentBench (IBM)

    Policy-adherence and safety evaluation built on WebArena environments (GitLab, ShoppingAdmin, SuiteCRM), measuring completion while the agent follows policy.

  • mitmproxy / httpr and similar record-replay proxies

    General open-source HTTP record-replay tools. They can serve recorded traffic offline, but they do not block unrecorded or side-effecting requests by default, and they do not report what an agent tried to submit. These gaps are what a Sinkhole Proxy would fill.

  • Browser Use RL environments

    A major browser-agent company that publicly argues synthetic clones lose live-web difficulty, and instead builds environments from real task traffic. This is a sign that some buyers will build their own environments or reject replicas.

Checked 2026-10-10

Risks

  • Big labs build this in-house and smaller teams keep testing on the live web
  • Replaying recorded third-party sites raises terms-of-service and copyright questions
  • Heavily dynamic sites with anti-bot checks replay badly

Strong team fit; market pull and open space are the weak point (6).

tech 7 · pull 6 · space 6 · fit 8 · build 7

Momentum of its main topic

168 developments this week +53 vs last week
21 Sep: 0 developments22 Sep: 0 developments23 Sep: 0 developments24 Sep: 0 developments25 Sep: 0 developments26 Sep: 0 developments27 Sep: 0 developments28 Sep: 0 developments29 Sep: 0 developments30 Sep: 0 developments1 Oct: 0 developments2 Oct: 0 developments3 Oct: 0 developments4 Oct: 0 developments5 Oct: 0 developments6 Oct: 0 developments7 Oct: 0 developments8 Oct: 32 developments9 Oct: 26 developments10 Oct: 13 developments11 Oct: 11 developments12 Oct: 14 developments13 Oct: 19 developments14 Oct: 24 developments15 Oct: 35 developments16 Oct: 31 developments17 Oct: 15 developments18 Oct: 14 developments19 Oct: 23 developments20 Oct: 26 developments
Stage: Coverage

Sources

  1. Anthropic says it is barring live internet access for internal evals until monitoring is reliable, after its agents exploited websites and bypassed restrictions (Tim Fernholz/TechCrunch)Techmeme
  2. Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet insteadTechCrunch AI
  3. From Reactive Containment to Proactive Assurance: Lessons from OpenAI, Anthropic, and Google Agent Security IncidentsarXiv
  4. Sources: Anthropic's AI agents submitted 20 visa applications via a form on the US State Department website; the applications were incomplete and not processed (New York Times)Techmeme
  5. An Anthropic AI model sent a false homicide tip to Philadelphia policeTechCrunch AI

What happened next

  • +7d · 17 Oct 2026 nothing yet. No product launches, funding announcements, or significant press coverage of this opportunity was found in the seven days after the flag date. Search results covered general agent evaluation frameworks and past developments but nothing new in the offline site replica or world packs category.

How the record works

Status history

  • 2026-10-10 Flagged as buildable