IdeaWell

Project ideas from Hacker News discussions.

Anthropic Is Building a Predictive Surveillance System to Monitor Activists

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

Four prevalent themes in the discussion

  1. AI used for surveillance and protest monitoring to protect corporate interests
  2. “Palantir's Maven system integrated Claude as its model for determining targets to bomb in Iran.” – culi
  3. “Anthropic contracts with a security agency … monitoring activists … that extra hour was critical… Without it our executives would have departed their meetings, they would have ran right into the heart of the disruption.” – cldellow

  4. “When next government become worse than the previous one, such technologies become handy for new emperor to keep power and crush dissent.” – varispeed

  5. Skepticism of the “good guy”/safe‑AI narrative; accusations of hypocrisy

  6. “Imagine defending a corporation because they were 'different' because they were here to deliver 'safe' and 'responsible' AI and are the 'good guys'.” – rvz

  7. “Remember guys: Anthropic are the good guys! They care about morals, alignment, and safety! So this invasive surveillance to crush dissent must be good for us!” – palmotea (sarcastic)

  8. Fear that AI enables authoritarian control, thought‑crime, and dystopian outcomes

  9. “AI is an authoritarian's dream.” – palmotea
  10. “What today is used for good, tomorrow can be used for evil.” – varispeed

  11. “Future is bright because it's going up in flames.” – 0c3ca83

  12. Distrust and power imbalance between ordinary people and tech elites; AI used to shield executives from public scrutiny

  13. “I think it is fundamentally a crisis of trust … I think that ordinary people don’t trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over.” – slowin (quoting Dario Amodei)
  14. “I'm being surveilled and profiled by these companies. Let their CEOs feel maximum heat.” – laimewhisps
  15. “Shielding the executives from the voices of the people… Critics can ask them anything, sure, but only on their terms.” – emsign

🚀 Project Ideas

PromptGuard: Local AI API Privacy Proxy

Summary

  • A self-hosted proxy that sits between users and AI API endpoints, stripping identifying metadata, logging requests locally, and allowing users to opt‑in or out of any data sharing with the provider.
  • Core value proposition: guarantees that your prompts and conversation history never leave your control unless you explicitly permit it, countering fears of covert reporting by AI companies.

Details

Key Value
Target Audience Privacy‑conscious developers, activists, journalists, and anyone using hosted AI APIs who fears surveillance or unwanted reporting.
Core Feature Intercepts HTTP(S) traffic to AI providers, removes headers/IPs, stores encrypted logs locally, and provides a toggle to block or allow transmission of prompts to the upstream API.
Tech Stack Go or Rust for the proxy core, MITM‑TLS via a local CA, SQLite + AES‑256 for log storage, optional Web UI in React/Vite.
Difficulty Medium
Monetization Hobby (open‑source); optional paid hosted instance or support contracts.

Notes

  • HN users lamented “closed AI is an authoritarian's dream” and wished they could “self‑host a powerful LLM” to avoid monitoring (e.g., SoftTalker, ryandrake). PromptGuard gives that control without sacrificing access to frontier models.
  • Enables discussion about what data is actually being sent; users can verify logs and detect unexpected exfiltration, addressing the trust crisis highlighted by Dario Amodei’s comments.

AIWatchdog: Browser Extension for AI Usage Auditing

Summary

  • A lightweight extension for Chrome/Firefox that monitors outbound requests to known AI APIs (Anthropic, OpenAI, etc.), logs prompts and responses locally, and alerts users when language matches patterns that could trigger internal reporting (e.g., threats, illicit content).
  • Core value proposition: gives users visibility into what they are sending to AI services and warns them of potential reporting risks before data leaves their browser.

Details

Key Value
Target Audience Everyday AI chatbot users, researchers, and privacy advocates who want to know exactly what their conversations contain and whether they might be flagged.
Core Feature Real‑time inspection of POST bodies to AI endpoints, keyword/heuristic scanning for risky content, local encrypted log, and optional push‑notification alerts.
Tech Stack JavaScript/TypeScript, WebExtension APIs, IndexedDB + CryptoJS for local encryption, optional background service worker.
Difficulty Low
Monetization Hobby (open‑source); possible premium features like advanced pattern rules or cloud sync as a subscription.

Notes

  • Commenters expressed fear that “if you type something threatening, the AI will report you to police” (cildellow, alistairSH). AIWatchdog makes that process transparent, letting users see exactly what triggers such actions.
  • Provides fodder for HN debates about the efficacy of keyword‑based moderation vs. semantic understanding, and can spark discussion on improving AI safety without compromising privacy.

PrivateInference: Verifiable Confidential LLM Inference via TEEs

Summary

  • A service that runs open‑source LLMs inside hardware‑based trusted execution environments (e.g., AMD SEV‑SNV, Intel TDX, AWS Nitro) and provides remote attestation proofs that no input data can be seen by the operator or leaked outside the enclave.
  • Core value proposition: lets users enjoy powerful AI models with cryptographic assurance that their prompts remain private, directly addressing the distrust of “closed AI” and fears of government‑mandated backdoors.

Details

Key Value
Target Audience Enterprises, developers, and privacy‑sensitive users who need GPU‑accelerated LLM inference but require provable data confidentiality (e.g., handling proprietary code, whistleblower communications).
Core Feature Deploy LLMs in TEEs, expose a standard OpenAI‑compatible API, and furnish attestation reports (via Intel IAS or AMD’s SNP verification) that users can verify before sending any data.
Tech Stack Rust/Go for attestation wrapper, Docker/Kubernetes for orchestration, GPU drivers, TEE SDKs, optional gRPC/REST API; frontend in Svelte for attestation verification UI.
Difficulty High
Monetization Revenue‑ready: usage‑based pricing (per token) with a free tier for low‑volume attestation verification.

Notes

  • Many commenters advocated for self‑hosting to avoid surveillance (e.g., “If you can afford to self‑host a powerful LLM, you will have more liberties” – cldellow). PrivateInference offers a middle ground: no need to manage hardware, yet with verifiable privacy.
  • Would likely generate extensive HN discussion about the practicality of TEEs for AI, reminiscent of debates on secure enclaves for blockchain or confidential computing.

AITransparencyLog: Community‑Curated Registry of AI Surveillance Incidents

Summary

  • A public, moderated log where users can submit verifiable reports of AI providers allegedly reporting users to authorities, sharing data with governments, or using prompts for surveillance. Each entry includes evidence (screenshots, timestamps, canary test results) and a confidence rating.
  • Core value proposition: creates a crowd‑sourced watchdog that increases transparency, deters covert data sharing, and empowers users to make informed choices about which AI services to trust.

Details

Key Value
Target Audience Privacy activists, journalists, researchers, and general users who want to hold AI companies accountable for potential misuse of their data.
Core Feature Submission workflow with evidence upload, automated canary‑test integration (users can run a standard test prompt that would trigger reporting if the provider logs it), moderation queue, searchable/filterable entries, and API for third‑party tools to check provider reputation.
Tech Stack Backend: Node.js/Python with PostgreSQL; Frontend: SvelteKit or Next.js; Canary test runner: CLI tool in Rust/Go; optional AI‑assisted evidence verification using open‑source LLMs.
Difficulty Medium
Monetization Hobby (open‑source, donation‑based); possible premium API access for enterprise risk‑monitoring teams.

Notes

  • HN thread repeatedly cited examples where “Anthropic reported the user to the police” and feared a slide into thought‑crime policing (cildellow, rottencupcakes). A transparent log would let the community corroborate or refute such claims, directly addressing the crisis of trust.
  • Encourages practical utility: developers could integrate the log’s API into their apps to warn users before they engage with a flagged provider, turning discussion into actionable tooling.

Read Later