Firefox Hands Mobile Users a Simple Off Switch for AI Features

Firefox just brought its AI controls to phones. With the release of version 151 on May 19, users on Android and iOS gained a straightforward way to shut down generative AI tools across the browser. One tap disables everything. Or they can select features one by one. The move extends a desktop capability launched earlier this year and sets Mozilla apart at a moment when many tech giants push AI defaults that users must hunt to disable.

The timing feels pointed. On the same day Google unveiled a wave of new AI capabilities at its I/O event, Mozilla delivered something rarer: choice. As reported by Engadget, the update arrives amid broader industry conversations about AI fatigue. Not everyone wants the browser suggesting tab groups, summarizing links, or generating image descriptions. Firefox now lets them say no. Cleanly.

But the story runs deeper than a settings toggle. Mozilla has spent years positioning itself as a defender of user agency in an AI-saturated web. Its work on open-source safety tools reveals both progress and persistent holes. In November 2025 the Mozilla AI team released benchmarks testing guardrails against agentic threats. They built a framework called any-guardrail to evaluate models on prompt injection and function-calling failures. Results showed promise in one area. Gaps elsewhere.

PIGuard stood out. The model achieved strong F1 scores detecting indirect prompt injections in email and tabular data pulled from the BIPIA dataset. It balanced precision and recall better than several alternatives. Yet when the same team turned to judging malformed function calls, performance collapsed. Custom judges like FlowJudge and GLIDER posted F1 scores below 0.5 even with few-shot examples. Explanations proved inconsistent. “Some models show promise for prompt injection detection, but critical gaps remain in protecting function-calling operations,” the researchers wrote. The finding matters for browser makers. As Firefox adds AI sidebar chatbots and on-device processing, these agent-like behaviors could expose new risks.

So far Mozilla has kept many of its AI experiments optional. The mobile controls mirror what arrived on desktop with Firefox 148 in February. A dedicated section in settings offers the master “Block AI enhancements” switch. Flip it and pop-ups, reminders, and new generative features disappear. Users who want translation, PDF alt text, smarter tab organization, link summaries, or a sidebar chatbot powered by Claude, ChatGPT, Copilot, Gemini or Mistral can enable them individually. Preferences persist across updates. No nagging.

“AI is changing the web, and people want very different things from it,” Mozilla stated in its February announcement. “We’ve heard from many who want nothing to do with AI. We’ve also heard from others who want AI tools that are genuinely useful.” The organization followed through. That consistency stands in contrast to competitors who bury opt-outs or tie AI to core functionality.

Voice search and page translation appear among the features now controllable on mobile. Some rolled out with version 151. Others were already present. The controls give users a single place to audit and adjust. For privacy-conscious professionals who manage fleets of devices or advise enterprises, the change removes friction. No more digging through scattered menus. No more worrying that an update re-enables unwanted data flows to third-party models.

Mozilla’s broader AI posture informs these choices. The organization collaborates with Anthropic on security research. In early 2026 it used Claude models to surface hundreds of Firefox vulnerabilities before they reached production. That work highlights a two-sided reality. AI can strengthen defenses. It can also introduce fresh attack surfaces if left unchecked. Browser-based AI agents that browse, fill forms, or call APIs on a user’s behalf demand careful constraints. The guardrail benchmarks Mozilla published last fall underscore how hard that remains.

Function calling evaluation proved especially stubborn. Models trained primarily on input and output moderation struggled to assess whether an agent invoked the right tool with the right parameters. Recall stayed low. False positives multiplied. The team concluded that current open-source options fall short for production agent safety. They open-sourced the testing harness anyway. Developers can now swap models and run their own experiments without rewriting evaluation code. The any-guardrail library supports both discriminative and generative judges. It abstracts away preprocessing differences. Practical progress born from acknowledged limits.

Industry observers note the difference in philosophy. While some browsers integrate AI deeply into search, recommendations, and even page rendering, Firefox treats these capabilities as add-ons. The mobile rollout reinforces that stance. Users who prefer a lean, non-AI browser experience can have one. Those who experiment with local translation or voice input keep the door open. Control stays with the person holding the device.

That matters for enterprise adoption. Security teams evaluating browsers for regulated environments often cite AI data leakage as a top concern. Summarization features that send page content to remote models create audit headaches. A blanket off switch simplifies compliance. It also signals to developers that Mozilla prioritizes configurability over forced engagement. In a market where retention metrics sometimes drive product decisions, this approach looks deliberate.

Of course challenges persist. On-device models require storage and battery. Cloud-connected chatbots raise questions about prompt logging even when providers promise privacy. Mozilla has emphasized that many translation and accessibility features can run locally with small models downloaded on first use. Yet the sidebar chatbot options connect to external services. The controls let users weigh those trade-offs themselves.

Recent coverage highlights the contrast with Google’s approach. Multiple outlets reported users struggling to mute AI overviews, Gemini integrations, and other defaults across Search, Android, and Chrome. Firefox’s binary toggle stands as a simpler alternative. One setting. Immediate effect. No hidden submenus. The difference may appeal to IT administrators tired of writing group policies to suppress AI nags.

Mozilla continues to expand its AI safety research. The November benchmarks focused on agent threats because future browser assistants could act with greater autonomy. Detecting prompt injection protects against manipulated web pages that trick chatbots into revealing data or taking unwanted actions. Reliable function-call validation would guard against incorrect tool use that might expose private tabs or bypass user intent. The current shortcomings suggest more work lies ahead. But the public release of testing tools invites the wider community to close those gaps.

For now the message from Firefox remains consistent. AI belongs in the browser only if the user invites it. On mobile that invitation comes with clear exit ramps. The feature arrives at a time when browser choice feels more meaningful than ever. As AI capabilities multiply across every app and website, the ability to step back carries real value.

Professionals who build, secure, or simply use the modern web will watch how this evolves. Will other vendors follow Mozilla’s lead and offer comparable one-tap controls? Or will AI features become so embedded that opting out feels like running an outdated product? Firefox has staked its position. The rest of the industry gets to respond.