Lumina

Vintage 2026-09-15

What changed, and how Lumina stays current.

IC3 2025 is brutal. “Good swarm” is not the unused button.

1,008,597 complaints. $20.877 billion reported. Cyber-enabled fraud about 85% of those dollars. People 60+ took a huge share. Crypto-linked: 181,565 complaints / $11.366 billion. AI-related as a new breakout: 22,364 / about $893 million. “99%+ unaddressed” is directionally right for individual cases; it is not a clearance rate IC3 prints. The Recovery Asset Team froze $679 million on 3,900 incidents — a 58% hit on that sliver, and still tiny against $20.9 billion.

source
FBI IC3 2025 Annual Report

1,008,597 complaints and $20.877 billion in reported losses. Cyber-enabled fraud ~85% of losses. Crypto-linked: 181,565 complaints / $11.366 billion. AI-related breakout: 22,364 complaints / ~$893 million. These are complaints, not clearance.

source
IC3 Recovery Asset Team / Financial Fraud Kill Chain (2025)

3,900 incidents, $1.164 billion in attempted theft, $679 million frozen (~58% on that sliver). Useful and still tiny against $20.9B. Most complaints are intake and pattern analysis, not “agent assigned, victim made whole.”

source
IC3 leadership capture-rate estimate (~10–12%)

FBI IC3 leadership (including former IC3 Director Donna Gregory) has estimated the center captures about 10–12% of victims — an order-of-magnitude undercount. This is not a table in the 2025 IC3 annual report.

Three planes. Never one agent with internet root.

  1. Plane A — wallet and wire. First-payee holds, mule graphs, freeze-as-default, crypto off-ramp tripwires. Banks, FinCEN 314(b), exchanges. Lumina cannot freeze a Chase wire. We teach you to move fast and file.
  2. Plane B — this product. Local guardian with a nutrition-label permission set, evidence packs, scam-friction, no outbound hunting. If the phone is already owned, Plane B on THAT phone is theater — use Witness.
  3. Plane C — government analysis factory. CJIS / FedRAMP High, IC3 corpus, ranked disruption packets, humans with badges. CISA JCDC, FBI strategy, Microsoft DCU-style civil process. Not a volunteer botnet of home PCs.
    source
    CISA Joint Cyber Defense Collaborative

    Voluntary sharing among vendors, cloud, FBI, NSA, DOJ. Coordination and incident sharing — not a consumer swarm and not a deputized home PC.

    source
    18 U.S.C. § 1030 (Computer Fraud and Abuse Act)

    Unauthorized access is a crime for private parties. § 1030(f) carves authorized law-enforcement activity. A consumer app disclaimer is not a warrant. Not legal advice.

Amodei vs Huang vs this app

A warning that a more capable misaligned agent swarm might, on a 6–12 month horizon, build persistent botnets is not a description of Grok, Claude, ChatGPT, or Gemini as they exist today. Those systems are boxed: no unconstrained persistence, no license to scan and seize. The stack that would vacuum crime at internet scale is close to the stack that can own the network. Nvidia-style “cybersecurity is the next AI boom” is enterprise SOC — continuous inference for patching and detection. It is not “catch the compound draining grandma.” If labs only chase that loop, chips boom and IC3 still prints a record year.

Street-level AI

Voice clones, fake “your cameras are down” emails, and cheap vision models make social engineering ordinary. The defense is boring: out-of-band verify, passkeys, second-channel alerts. The tutor will not click the link for you.

Optics

Lenses in the 1–2 mm class are documented in academic hidden-camera work. Multi-angle LED reflection research exists. Your phone is a helper.

source
LAPD (KAIST) — smartphone ToF hidden-camera research

Academic result: tiny lenses may be only 1–2 mm; unaided visual search performed poorly compared with sensor-aided methods. Phones are helpers, not magic scanners.

source
Hide-and-Sweep / SweepLED (ACM, 2025/26)

Multi-angle LED illumination can distinguish camera-lens reflections from ordinary shine — still research-grade, not a guarantee.

Tiny device scale

Update loop

  1. Curriculum ships in app updates (this repo). Intel page shows the vintage.
  2. Gemini model ID is an env var so deprecations don’t require a redesign.
  3. Tutor is grounded in modules + guardrails, not a free-range search agent.
  4. Local heartbeats and SHA-256 seals are clocks on this device. Public stamps belong on a helper’s machine when this one is hostile.
    source
    OpenTimestamps

    Public, append-only timestamp of a SHA-256. Proves a bytestring existed at time T. Does not prove the bytes are true, and does not hide the fact that you published.

  5. Sources on stats are tooltips, not vibes.

Vehicle envelope

Vehicle inspection points