Anthropic has begun rolling out “Claude for Chrome”, a browser extension that lets its Claude AI system observe a user’s web activity and perform tasks such as navigating sites, filling out forms and drafting emails. The company is releasing the tool as a research preview to 1,000 customers on its Max subscription tier, which costs between $100 and $200 a month, and has opened a wait-list for wider access. To curb misuse, Anthropic says the agent requests explicit permission before high-risk actions such as purchases or publishing content. Users can blacklist specific domains, while the company has pre-emptively blocked access to financial, adult and piracy sites. Early internal tests indicate that a suite of defences has lowered successful prompt-injection attacks to 11.2% from 23.6%. The move intensifies a race to embed AI directly in the browser. Perplexity’s Comet launched earlier this month, Google has been weaving Gemini into Chrome, and OpenAI is reportedly preparing a competing agent. Anthropic’s tight cap on testers is meant to gather real-world feedback before a broader commercial release. Separately, Anthropic published an August threat-intelligence report describing attempts to weaponise Claude for cybercrime. One attacker used the model to automate data-extortion operations against at least 17 healthcare, government and other organisations, demanding up to $500,000 in bitcoin. The company says it detected and blocked the accounts, tightened filters and developed new classifiers to spot similar abuse.