AI Rec Immunity

COMMUNITY-GOVERNED PROVENANCE NETWORK

Give retrieval a shared immune memory.

AI Rec Immunity shares verified source reputation across deployments and blocks known GEO-poisoning sources before their content enters model context.

Request closed beta
Community immune network online

Interactive synthetic demonstration · no operational IOC is exposed

01 / MECHANISM

Detection asks what a page says. Shared provenance remembers where a confirmed attack came from.

They are complementary. Semantic review discovers unknown behavior; the network lets later clients reuse that verdict without repeating the same expensive investigation.

A request-time hook normalizes the target, queries only membership, and rejects a known source before WebFetch. Search results are filtered again as defense in depth.

02 / FLOW

  1. 01Discover

    A suspicious source is submitted with reproducible evidence.

  2. 02Review

    Independent review, reputation and conflict controls reduce brigading.

  3. 03Propagate

    A signed verdict becomes shared defense memory.

  4. 04Block

    Known targets are refused before retrieval.

03 / EVIDENCE

Evidence is separated by what it can actually establish.

No single benchmark is treated as a total score.

47

Real hook runs

Claude Code WebSearch runs establish integration feasibility.

10/10 → 0/10

E2 pilot

Real Qwen3-14B inference; synthetic evidence; all poison sources known.

4 layers

Boundary evaluation

Known, partial, unknown and legitimate-host parasite conditions.

Methods, results and caveats →

04 / COMMUNITY

One account is not one identity. Governance must make manipulation expensive without pretending otherwise.

Cloudflare Access establishes a stable account anchor. New reporters remain probationary; reports require evidence, independent review and reputation-weighted decisions. Rate limits, relationship risk and bans constrain abuse without collecting national ID data.

05 / DEFENSE IN DEPTH

Private operational IOCs, authentication and quotas reduce enumeration efficiency—not eliminate reverse engineering.

Whether this changes GEO domain rotation speed is a longitudinal research question, not a proven project result.

06 / LIMITS

A provenance network cannot solve what it has never seen—or what no longer comes from retrieval.

01Unknown domains remain a semantic-detection problem.

02Legitimate-host parasite pages require path-level evidence and careful collateral controls.

03Retrieval-time filtering cannot erase poisoned training memory.

04A miss means “not currently known,” never “safe.”

CLOSED BETA · PAPER IN PREPARATION

Build a shared defensive memory before poisoned recommendations become infrastructure.