AI privacy and safety · How Founders Use AI
Is it safe to let an AI agent or AI browser log in to my email, bank or shopping accounts?
The specific risk is prompt injection: instructions hidden in a web page, email or document tell the agent to do something you never asked, like forwarding your emails or entering your card details. Brave showed Perplexity's Comet browser could be tricked this way in 2025, and the makers say it is not solved: Anthropic reports 0.3% of strong attacks still getting through against Claude Fable 5 in Chrome with all its safeguards on. So give an agent only the accounts a task needs, keep it off banking unless you watch every step, ask for narrow jobs ('reply to these three emails', not 'handle my inbox'), and read each confirmation before approving. ChatGPT's Lockdown Mode (Settings, Security) turns off browsing, agent mode and connectors when you work with sensitive files.
On the AI map, 29 October 2026 AI assistantsAI browsersAvatars and voice
Go deeper9 sources
-
Understanding prompt injections: a frontier security challenge
OpenAI · Article · Nov 2025
A plain explanation of prompt injection with examples, and the habits that reduce it (its Atlas examples refer to a browser OpenAI has since retired).
-
Introducing Lockdown Mode and Elevated Risk labels in ChatGPT
OpenAI · Article · Feb 2026
The setting, open to personal accounts since June 2026, that switches off web access, agent mode and connectors.
-
Claude in Chrome is generally available
Anthropic · Article · Aug 2026
How Claude's browser agent checks each action, its measured attack success rates, and how to go back to approving every step.
-
Mitigating the risk of prompt injections in browser use
Anthropic · Article · Nov 2025
Anthropic's own walk-through of a hidden-instruction attack on an email agent, and why 1% success is still real risk.
-
Agentic Browser Security: Indirect Prompt Injection in Perplexity Comet
Brave · Article · Aug 2025
Security researchers show how a web page could steer an AI browser into handing over a user's account details.
-
The lethal trifecta for AI agents: private data, untrusted content, and external communication
Simon Willison · Article · Jun 2025
The clearest rule of thumb: an agent with your private data, outside content and a way to send messages can be turned against you.
-
2026: The State of Consumer AI
Menlo Ventures · Article · Sep 2026
How many people already give agents their email (36%) and financial accounts (20%), and the 32% who let AI act without a final sign-off.
All 9 sources
-
What Is Prompt Injection? The Real Risk in AI Agents
KodeKloud · Video · Jul 2026
A short video showing how hidden instructions hijack an agent.
-
Tokens in the Stream - Why LLMs are inherently insecure and prompt injection will persist
Security Now (Steve Gibson, TWiT) · Podcast · Aug 2026
A long-running security show on why prompt injection is built into how these models read text.
Listen here