Restricting AI agents in mobile operating systems platforms from capturing data
Signal’s move to block OS-level AI agents from reading content on its encrypted messaging app is prompting broader questions about what “privacy” realistically means on modern phones. Commenters largely welcome efforts to resist screen-scraping assistants from Apple and Google, but point out that messages can still leak via recipients’ devices, OS accessibility features, cloud backups and push-notification metadata. Many also criticize Signal’s reliance on phone numbers or KYC-linked payments, centralized infrastructure and Intel SGX, arguing that these choices limit anonymity and create legal and technical choke points compared with more decentralized or self-hosted alternatives.
AI Agents and Screen Access
- Core issue: OS-level AI agents (Siri, Android “on-screen” AI, accessibility services) can see rendered text, notifications, and screenshots from Signal, bypassing E2EE and at-rest encryption.
- Examples: past iOS behavior where invoking Siri could send a screenshot of an open Signal chat to Apple’s cloud; Android accessibility APIs that can read any text control.
- Some think OSs should add explicit “AI/agent access” permissions like location, ideally with per-app controls and “always ask/never allow” options.
- Others note that full protection is impossible: any participant can copy, screenshot, or photograph messages.
User Choice vs Counterparty Privacy
- Some users want OS/AI access for indexing or convenience (e.g., Siri/Spotlight search over Signal chats) and resent Signal blocking it by default.
- Counter-argument: your contacts may not want their messages fed to cloud AIs; their privacy interests justify restrictive defaults.
- Suggestions: per-conversation modes (e.g., “secret” chats tied to disappearing messages) where anti-AI protections are enforced for all participants.
Signal’s Privacy Model and Trust Debates
- Supportive view: Signal offers strong content privacy via audited, open-source clients, formally analyzed protocols, reproducible builds, and a record of providing almost no data in legal requests. Metadata collection is minimized, though not eliminated.
- Critical view: US jurisdiction, centralized servers on AWS, and Intel SGX-based features create opportunities for metadata access by state actors.
- Some allege Signal now stores profiles and contact/social-graph data (and optionally message backups) in the cloud, protected only by a PIN and vulnerable SGX; they argue the published privacy policy is outdated or misleading. Others respond that these data are encrypted and that no mass-market app does better.
Identity, Spam, and Payments
- Long-standing complaint: requiring a phone number undermines anonymity.
- Thread notes an Android beta for phone-numberless registration, but it relies on KYC-linked app-store payments; critics say this just swaps one identifier for another.
- Defenders argue phone numbers or paid registration are needed to deter spam/bot floods; some want anonymous crypto (e.g., Monero) as an alternative.
Centralization, Federation, and Alternatives
- Critiques: no federation, no public API, discouraging third-party clients, single global infrastructure and key; this centralization is seen as structurally incompatible with maximal privacy/anonymity.
- Others say federation and self-hosting (Matrix, XMPP, SimpleX, Jami) are better for high-threat users but much worse for usability and network effects. Signal is framed as “best practical option for the masses,” not an anonymity tool.
OSes, Attestation, and Devices
- Some argue real privacy requires trusted OSes (GrapheneOS, hardened Linux) and user control over attestation and screen-capture.
- Others warn that remote attestation and DRM-like measures, if used by apps like Signal, can become user-hostile by treating the device owner as part of the threat model.
Usability and Client Implementation
- Complaints about Signal Desktop: Electron bloat, memory use, weak at-rest protection (DB encrypted but key stored alongside), and reduced screenshot capability on some platforms.
- Users also report friction with VPNs (CAPTCHA loops) and missing QoL compared to WhatsApp/other messengers.
- Some accept these as trade-offs given Signal’s small team and cross-platform burden; others see them as signs of a “careless” or under-resourced project.