Merge pull request #11 from saquibreja7-hash/feature/ai-image-checker

AI image provenance checker + survivor-security roadmap

Author: saquibreja7-hash

asmita/.env.example

@@ -31,6 +31,18 @@ RESEND_FROM_EMAIL="Asmita <Notice@meriasmita.org>"
 # Resend webhook signing secret (starts with whsec_). Get from Resend dashboard → Webhooks.
 RESEND_WEBHOOK_SECRET=
 
+# --- Content provenance (OpenAI) --------------------------------------------
+# Powers the optional "deeper check" on /check-image: reads C2PA credentials and
+# the SynthID watermark. The image is forwarded to OpenAI in memory and never
+# stored by us (see docs/adr/002-consented-encrypted-media.md).
+#
+# OFF-BY-DEFAULT SAFETY GATE. The /api/check-image route returns 404 and the UI
+# hides the deeper check unless ENABLE_PROVENANCE_CHECK=true. This forwards media
+# to a third party, which ADR 002 forbids until legal sign-off, so do NOT set the
+# flag in production or preview until those blockers clear. Local dev only for now.
+ENABLE_PROVENANCE_CHECK=
+OPENAI_API_KEY=
+
 # --- Queue + cache ----------------------------------------------------------
 # Vercel serverless functions are stateless: REDIS_URL must point to a
 # managed Redis (Upstash, Azure Cache for Redis) for production.

asmita/docs/adr/002-consented-encrypted-media.md

@@ -0,0 +1,60 @@
+# ADR 002: Media May Be Stored, But Only Encrypted, Consented, and Time-Limited
+
+Status: proposed (blocked on legal sign-off)
+
+## Context
+
+ADR 001 and the original safety charter state that the server never fetches, stores, or processes intimate image content. Only PDQ hashes and text reach the server. This kept Asmita architecturally simple and legally clean.
+
+The survivor-security platform direction (see `docs/product/survivor-security-roadmap.md`) needs capabilities that this rule forbids:
+
+- An evidence vault that preserves screenshots, threat messages, and payment demands so a survivor can escalate to a platform, the police, or a court after the original content is deleted.
+- Optional server-side content-authenticity assessment (deepfake detection) for images that carry no client-readable credential.
+- Opt-in resurfacing monitoring.
+
+The decision recorded here is to change the charter from "no media, ever" to "media only under strict, consented, encrypted, time-limited conditions."
+
+## Decision
+
+We will allow media storage and server-side media processing under all of the following conditions. Every one is a hard requirement, not a guideline.
+
+1. **Adults only.** Any path that could involve a person under 18, or an image created when the subject was under 18, is routed to the child-protection flow and is never uploaded to Asmita. This preserves the existing "minors never reach the case flow" principle without exception.
+2. **Explicit, specific consent** per feature. Consent to open a case is not consent to store an image. Consent to store evidence is not consent to run detection. Each is asked separately and recorded in the audit log.
+3. **Encrypted at rest**, with keys managed so that ordinary staff and general infrastructure access cannot read the plaintext.
+4. **Time-limited by default.** Every stored object has a deletion date. The survivor can delete immediately at any time. Deletion is real, not a soft flag.
+5. **In-memory for detection.** Server-side authenticity checks process the image in memory and discard it. No detection path writes the image to disk or database.
+6. **Purpose-bound.** An object stored for evidence is used only for that survivor's case. It is never reused for training, aggregate research, or identifying other people.
+7. **Never a gate.** The result of any detection never decides whether a survivor receives help. See the roadmap's content-authenticity section.
+
+## Enforcement (to be built alongside the features)
+
+- The "no media, ever" lint tests and `no-fetch` monitor remain in force for the takedown and notice pipelines. Those pipelines still never touch media. The new media paths are separate, opt-in, and covered by their own tests.
+- A minor-exclusion gate must sit in front of every upload entry point, tested to fail closed.
+- Retention and deletion must have automated tests proving objects expire and hard-delete.
+
+## Documented exception: OpenAI content-provenance check
+
+The `/check-image` deeper check (`/api/check-image`) forwards the image to OpenAI's `content_provenance_checks` endpoint to read C2PA credentials and the SynthID watermark. This is a third-party processor, which conditions 5 and 6 above would otherwise forbid.
+
+It is allowed as a narrow, opt-in exception because:
+
+- It is **opt-in per image**, behind an explicit consent line naming OpenAI, on top of the adults-only gate.
+- The image is **held only in memory** on our server for the request and **never written down** by us.
+- It returns a **fact-based signal** (a credential or watermark), not a guess, so it cannot falsely accuse a genuine photo, and the result **never gates help**.
+- It is an interim measure until direct SynthID access or our own in-house model is available.
+
+Residual risk accepted by Saquib (2026-08-02): OpenAI receives the image and retains it briefly under its own API data policy. This exception must be revisited if we move detection fully in-house.
+
+Current shipping state: the route and its UI are behind `ENABLE_PROVENANCE_CHECK`, off by default. With the flag unset, `/api/check-image` returns 404 and the deeper-check UI is hidden, so setting `OPENAI_API_KEY` alone does not make the feature live. The flag stays off in production and preview until the blockers below clear.
+
+Known limitation to resolve before enabling: the adults-only gate is client self-attestation, which a caller can bypass on a public endpoint. Before the flag is turned on anywhere real, the route must be bound to an authenticated case that already passed the adult age gate, and the in-memory IP rate limit must move to the Redis driver (it does not hold across stateless serverless instances).
+
+## Status and blockers
+
+This ADR is **proposed, not accepted**. It must not ship until:
+
+- Indian counsel reviews it against the DPDP Act, the IT Rules, and the MeitY NCII SOP, including obligations around holding intimate imagery of adults.
+- The retention, encryption, and deletion design has a written security review.
+- The child-protection routing has an explicit protocol (see the POCSO/Take It Down work).
+
+Until then, features that depend on it are designed and specced but built behind a flag that is off in all environments, or built only up to the point where no real media is stored.

asmita/docs/product/survivor-security-roadmap.md

@@ -0,0 +1,119 @@
+# Meri Asmita: Survivor-Security Platform Roadmap
+
+Status: draft for review
+Owner: Saquib
+Last updated: 2026-08-02
+
+## The shift
+
+Meri Asmita moves from a takedown service to a single, trauma-informed response centre. A survivor should be able to understand what is happening, preserve evidence, reduce further spread, submit and track removals, escalate, secure their accounts and devices, and reach legal, psychological, and practical support.
+
+The organising promise is not "we can prove whether an image is AI-generated." It is:
+
+> We help you regain control, by preserving evidence, reducing distribution, securing your accounts, pursuing removals, and connecting you with the right support.
+
+**Deepfake detection is a supporting feature, never a gate.** Real NCII, face swaps, nudify images, mislabeled images, and threats to publish all get the same help. Whether an image is authentic, altered, or generated never decides whether we assist.
+
+## Where we already are
+
+The current codebase covers a meaningful slice of Phase 1 already:
+
+- URL takedown, notice generation and dispatch (`notice-*.ts`, `case-ops.ts`).
+- Client-side PDQ hash submission (`hash-submission.ts`, `hash-dispatch.ts`).
+- POCSO / minor routing (`/minor-support`), keeping minors out of the adult case flow.
+- Admin dispatch with human contact verification and legal review gates.
+- English and Hindi throughout, OTP login, audit log, rate limiting, CSRF.
+- Client-side content-credential reader (`/check-image`, `ai-provenance.ts`), the first step of the authenticity feature.
+
+So this roadmap is mostly about adding the response-centre layer around a working takedown engine, not starting over.
+
+## The charter decision that governs everything
+
+Moving to an evidence vault, server-side detection, and monitoring breaks the original "no media, ever" rule. We have decided to change it to "media only under strict, consented, encrypted, time-limited conditions." That decision is written up in `docs/adr/002-consented-encrypted-media.md` and is **blocked on legal sign-off**.
+
+Practical consequence: everything in this roadmap that stores or server-processes media is designed now but does not ship until counsel and a security review clear ADR 002. Features that do not touch media can proceed in parallel.
+
+## Product structure
+
+Five areas, matching how a survivor thinks, not how the backend is organised:
+
+1. **Get Help Now** — triage for published content, threats, sextortion, deepfakes, impersonation, account compromise.
+2. **My Case** — status, URLs, reports, deadlines, messages.
+3. **Protect My Accounts** — account, device, privacy, impersonation checks.
+4. **Preserve Evidence** — secure evidence log and downloadable reports.
+5. **Support and Recovery** — legal aid, counselling, safety planning, resources.
+
+## Roadmap
+
+Each feature lists what it is, what it maps to in the code, and its main gate or risk. "Gate" means something that must be true before it ships.
+
+### Phase 1: Survivor response MVP
+
+Goal: turn the takedown form into a response centre. Most of this touches no media, so it can move without waiting on ADR 002.
+
+| # | Feature | Maps to / build | Gate |
+|---|---|---|---|
+| 1 | Guided "Get Help Now" triage | New `(public)` flow feeding case creation; decision tree, not a long form | None. Ships first. |
+| 2 | Safe access + Quick Exit | New global component: Quick Exit button, double-Escape, neutral tab title/favicon, session timeout, `noindex` on dashboards, no sensitive text in emails/notifications | None. Safety-critical, build early. |
+| 3 | Anonymous / pseudonymous case creation | Extend case model + auth: random case ID, password/passkey, optional contact, recovery code, three modes (self-help / anonymous assisted / verified legal) | Review data model against DPDP minimisation. |
+| 4 | Case dashboard | Extend `(victim)` dashboard into a clear timeline with per-URL status, references, escalation deadlines, no vague "processing" | None. |
+| 5 | URL + evidence collection (text only) | Extend URL intake to capture usernames, timestamps, threat text, payment demands as text/hashes. No image storage in Phase 1. | Keep to text + hashes to stay inside current charter. |
+| 6 | Platform-specific report generator | Extend notice generation with a maintained platform directory (reporting route, NCII policy, deepfake policy, India grievance officer, timelines, last-verified date) | Human contact verification stays required. |
+| 7 | Google / Bing de-index assistance | New notice type; treat de-indexing as separate from source removal | None. |
+| 8 | India NCRP / Grievance Officer / GAC pathways | New India NCII Action Centre flow; version-controlled deadline rules engine | Deadlines (2-hour vs 24-hour) reviewed by counsel before stated as fact. |
+| 9 | StopNCII + Take It Down routing | Guided handoff, not a rebuild: adults to StopNCII, under-18 to Take It Down | Correct minor routing, tested. |
+| 10 | Secure caseworker messaging | New secure inbox; role-based access; permissioned case summary so survivor never re-tells their story | Role-based access model reviewed. |
+| 11 | English + Hindi | Existing i18n | None. |
+| 12 | Explicit retention + deletion controls | Case withdrawal, delete, retention clock surfaced to survivor | Foundation for ADR 002. |
+
+### Phase 2: Security and support
+
+Goal: the "regain control" layer. Several items here store media and are gated on ADR 002.
+
+| # | Feature | Maps to / build | Gate |
+|---|---|---|---|
+| 7v | Encrypted evidence vault | Encrypted, time-limited object store; SHA-256, access log, redacted copy, downloadable evidence/legal packets | **ADR 002 + security review.** Minors hard-excluded. |
+| 12 | Account Security Checkup | Platform-specific checklists; asks first whether abuser has device access | None (content pages). |
+| 13 | Stalkerware / device safety assessment | Questionnaire + safer-options guidance; does not tell everyone to uninstall | Follow Coalition Against Stalkerware guidance. |
+| 11 | Sextortion emergency mode | Prominent "I am being blackmailed" path; urgent info first, do-not-pay guidance | None. High priority within Phase 2. |
+| 17/18 | Support directory + mental-health/grounding | Verified directory (One Stop Centres, DLSAs, cybercrime units, counselling, LGBTQ+, disability, child protection); grounding exercises and referral | Listings vetted, last-verified dates. |
+| 19 | "Help someone else" mode | Authorised-representative accounts with survivor-controlled, revocable permissions | Consent model; aligns with MeitY SOP representative provision. |
+| 20 | More Indian languages + low-bandwidth | Bengali, Urdu, Tamil, Telugu, Marathi, Malayalam; text-first, save-and-resume, screen-reader, code-mixed | Translation review. |
+
+### Phase 3: Proactive protection
+
+Goal: after partnerships and governance. Highest risk, most governance.
+
+| # | Feature | Maps to / build | Gate |
+|---|---|---|---|
+| 9 | StopNCII NGO / Global Clearing Centre participation | Partnership application; only run our own hash bank once we have authorised recipients | Partnership + legal. |
+| 10 | Resurfacing / exposure monitoring | Opt-in, survivor-controlled searches (hashes, name, known URLs, impersonation) | Consent + governance. |
+| 15 | Content-authenticity assessment (deepfake detection) | The `/check-image` reader plus optional server-side detector (our own model on Cloudflare Containers, in-memory, discarded). Strengthens reports, never gates help. Shows "likely," never "97% fake." | **ADR 002.** Framed as non-conclusive. |
+| 8 | Evidence-capture extension / bookmarklet | Captures pages the survivor already views; blurs explicit content; no covert surveillance | Security + abuse review. |
+| face/voice | Opt-in face / impersonation search | Separate explicit consent; encrypted biometric templates; human review of candidates, no auto-accusation; deletion controls | Consent design, legal review, accuracy testing, deletion controls. Do not launch without these. |
+
+## Internal tools (build alongside the phases they serve)
+
+- **Caseworker console**: risk level, survivor's requested outcome, URLs, deadline clock, actions, evidence refs, consent permissions, history, reappearance alerts.
+- **Reviewer-wellbeing protections**: blurred thumbnails by default, text-only where possible, click-and-hold to reveal, no autoplay, limited viewing time, access logging, two-person review for sensitive decisions.
+- **Audit and accountability**: who accessed each object and why, what was sent, which template and legal version, survivor authorisation, deletion events, detector version. Extends the existing audit log.
+
+## What we are deliberately not building
+
+- No public "upload and identify NCII" scanner. It becomes an abuse tool.
+- No private CSAM database. Minor cases go to authorised child-protection systems.
+- No promise of whole-internet removal. We state exactly which platforms, URLs, and search results were addressed.
+- No AI detection as a gatekeeper.
+- No broad face recognition without the full governance package.
+- No indefinite retention of intimate media.
+
+## Immediate next actions
+
+1. **Stop the Cloudflare deepfake-detection container work.** It is a Phase 3 supporting feature, not the centre. The client-side reader already shipped is enough until then.
+2. **Start counsel review of ADR 002.** Everything media-related in Phase 2 and 3 waits on it, so the clock should start now.
+3. **Build Phase 1, items 1 and 2 first** (triage front door + safe access / Quick Exit). They are low-risk, touch no media, and reframe the entire app immediately.
+4. Keep the existing "no media, ever" lint tests and no-fetch monitor in force for the takedown and notice pipelines. The new media paths are separate and get their own tests.
+
+## Sources
+
+Design references: StopNCII, Take It Down (NCMEC), eSafety Commissioner (AU), Revenge Porn Helpline (UK), CCRI Safety Center, Chayn/Bloom, NNEDV Safety Net, Coalition Against Stalkerware, Lila.help, MeitY NCII SOP, I4C/NCRP. Full links in the source brief that seeded this roadmap.

asmita/package-lock.json

@@ -31,6 +31,7 @@
     "bullmq": "^5.76.7",
     "docx": "^9.7.1",
     "dotenv": "^17.4.2",
+    "exifr": "^7.1.3",
     "ioredis": "^5.10.1",
     "jose": "^6.2.3",
     "next": "16.2.12",