Name the fight · public

Threat model

Privacy products fail when they market vibes. A threat model names who might hurt you, what they want, what we do, and what we still cannot stop.

ddl-threats/1 · updated 2026-07-11

  • T1Upstream search / model provider

    reduced
    Wants
    Link prompts or queries to a person or durable profile
    Asset
    Query text, prompt text, IP, account ids
    Mitigation
    Server-side proxy; no client IP to provider; no user-tied conversation ids; hashed aggregate stems only
    Residual
    Provider may log content; proxy operator sees content in transit
  • T2Curious proxy operator (us)

    reduced
    Wants
    Build dossiers from demo usage
    Asset
    Queries, prompts, identity
    Mitigation
    Schema forbids who→what joins; Glass Box + unit tests; invite-only reduces casual abuse
    Residual
    We could change the code; verifiability and public claims make silent betrayal harder to hide
  • T3Bot / scraper burning paid APIs

    mitigated
    Wants
    Drain Brave / model budget
    Asset
    API keys, money
    Mitigation
    Invite-only allowlist, HumanVerify, rate limits, constant-response OTP
    Residual
    Compromised allowlisted mailbox
  • T4Passive network observer on the client path

    reduced
    Wants
    See destinations and metadata
    Asset
    SNI, timing, destinations
    Mitigation
    TLS to our origin; third-party calls originate from server
    Residual
    Observer still sees you talk to duckduckllama.com
  • T5Malicious page under tracker analysis

    reduced
    Wants
    Attack the fetcher (SSRF, huge payloads)
    Asset
    Server resources, internal network
    Mitigation
    http(s) only, timeouts, size cap, no internal redirect follow into RFC1918 (best-effort)
    Residual
    SSRF hardening can always deepen
  • T6Future agent tooling

    reduced
    Wants
    Exfiltrate PII via tool calls
    Asset
    User content in agent loops
    Mitigation
    PII gateway + egress policy labels; preflight preview before send
    Residual
    Pattern detectors miss novel PII; full agent sandbox not shipped
  • T7Compelled disclosure / breach of our DB

    accepted
    Wants
    Historical content
    Asset
    Anything durable
    Mitigation
    Minimise durable content; no raw queries/prompts stored; short TTLs
    Residual
    Allowlisted emails and session hashes still exist

Machine-checkable proofs: /verify · interactive demo stays invite-only.