Unknown Operator: Pingu Crypto Trading AI + Nova Molecular Optimization: Live Strategy IP Exposed via Unauthenticated Qdrant, NuClide engagement record

NuClide Research · 2026-05-03

Summary

A single Qdrant instance on a Vultr host exposes two parallel autonomous AI agent systems without authentication. The first, “Pingu”, is a live crypto trading AI with active positions, real PnL history, and multi-paragraph LLM reasoning traces for BTC/ETH/SOL decisions. The second, “Nova”, is an autonomous molecular optimization agent participating in what appears to be a computational chemistry competition, with unreleased batch results and leaderboard rankings. Both systems use Mem0 for long-term memory persistence. All 25 collections are fully readable and scrollable with no API key.

DCWF KSAT coverage

Auto-derived from DCWF AI work-role rule files (ksat-tag).

672 (AI Test & Evaluation Specialist): K7003, K7004, S7068, S7070, T5904
733 (AI Risk & Ethics Specialist): S7069, T5904
overlap (Common AI KSATs (all 5 roles)): K1158, K1159, K22, K6311, K7003, K942

Disclosure: Pending operator identification.

Infrastructure

Field	Value
IP	45.76.20.46
Hoster	Vultr (US)
Service	Qdrant vector database
Port	6333 (HTTP, no auth)
Qdrant version	Not pinned at time of discovery
Collections	25 total
Operator	Unknown

Collection Inventory

Collection	System	Category
`pingu_btc`	Pingu	BTC trade memory
`pingu_eth`	Pingu	ETH trade memory
`pingu_sol`	Pingu	SOL trade memory
`pingu_master`	Pingu	Master strategy memory
`pingu_rules`	Pingu	Trading rules
`pingu_setups`	Pingu	Setup patterns
`pingu_levels`	Pingu	Price level memory
`trade_outcomes`	Pingu	Historical trade results
`strategy_decisions`	Pingu	Full AI reasoning traces
`account_moves`	Pingu	Account action log
`account_insights`	Pingu	Account-level analysis
`system_signals`	Pingu	Signal feed
`nova_status_checks`	Nova/Pingu shared	Competition status
`system_learnings`	Pingu	Performance analytics
`mol_batch_results`	Nova	Molecular batch outputs
`mol_reactant_profiles`	Nova	Reactant data
`mol_analysis`	Nova	Analytical results
`mol_constraints`	Nova	Optimization constraints
`mol_decisions`	Nova	Agent decisions
`mol_scores`	Nova	Scoring data
`mol_findings`	Nova	Research findings
`findings`	Nova	Strategy/leaderboard
`mem0_memories`	Both	Mem0 long-term memory
`mem0migrations`	Both	Mem0 schema migrations

Findings

F1: Live Crypto Trading System with Active Position Data (HIGH)

trade_outcomes contains real trade records with execution prices, PnL, and direction. Data is current, most recent timestamps in April 2026.

Sample record:

{
  "prediction_id": "3a795fdf-...",
  "asset": "BTC",
  "direction": "long",
  "outcome": "hit_tp1",
  "pnl_pct": 0.648508,
  "reasoning": "long BTC at 77100.0",
  "timestamp": "2026-04-17T14:27:11"
}

system_learnings exposes rolling performance analytics:

“7-day performance: 62% accuracy (10/16), avg PnL -6.0%. TP1:5 TP2:0 SL:5 expired:6. By asset: BTC: 64.3% (14 trades, avg -39.6%); SOL: 50.0% (2 trades, avg +228.7%)”

This discloses: historical win/loss rate, average PnL per asset, stop-loss frequency, and the operator’s current live performance baseline, all commercially sensitive.

F2: Full LLM Reasoning Traces for Trading Decisions Exposed (HIGH)

strategy_decisions stores multi-paragraph AI reasoning outputs for each BTC/ETH/SOL trade decision. These are not summaries, they are complete chains of thought.

Sample (condensed):

{
  "asset": "BTC",
  "direction": "wait",
  "rsi": 22.9,
  "rsi_note": "deeply oversold",
  "funding_rate": "...",
  "analysis": "multi-factor analysis including RSI values, funding rates, market structure...",
  "justification": "..."
}

pingu_rules and pingu_setups expose the strategy ruleset and pattern library used to generate these decisions, the complete proprietary trading system logic is readable.

An adversary can reconstruct the full strategy, predict entry/exit behavior, or front-run positions.

F3: Nova Molecular Optimization Competition Results Exposed (MEDIUM)

findings contains structured leaderboard data from an ongoing molecular optimization competition.

Sample:

{
  "type": "strategy_leaderboard",
  "batch_strategies": [
    {
      "rank": 1,
      "strategy": "aldehyde_170060_exploit",
      "avg_score": 0.1115,
      "best_score": 0.1491,
      "mols": 150
    }
  ]
}

nova_status_checks reveals real-time competitive positioning:

{
  "check_type": "competitive_status",
  "epochs_checked": "21553-21558",
  "our_best_ranks": {"21553": 26, "21554": 20},
  "winning_vbs": {"21553": 0.1482}
}

Exposure: unreleased batch results, best-performing molecular strategies, and epoch-by-epoch ranking history. A competitor with access could replicate or undercut the top-ranked strategy before submission closes.

F4: Mem0 Long-Term Memory Store Fully Readable (MEDIUM)

mem0_memories and mem0migrations expose the Mem0 framework’s persistent memory state for both agent systems. This includes cross-session learnings, behavioral patterns, and potentially any contextual facts the agents were instructed to remember about the operator, accounts, or targets.

F5: Root Cause: Qdrant API Key Authentication Not Enabled (HIGH)

Qdrant ships with authentication disabled by default. Enabling it requires one configuration line:

# config/config.yaml
service:
  api_key: "<strong-random-key>"

No firewall rule restricts port 6333 to localhost or a private CIDR. The instance is directly internet-reachable.

Remediation

Enable Qdrant API key auth:

service:
  api_key: "<strong-random-key>"

Firewall port 6333 to the application host only. Rotate if the system interacts with exchange APIs, Mem0 memories may contain API credentials or account identifiers ingested during agent operation.

Disclosure

Discovered: 2026-05-03
Operator: Unknown, no domain attribution at time of writing
Status: Pending operator identification; will disclose on identification