# Operation ArgonHunt: Open Directory Exposes Turkiye Crypto Exchange Breach and 44,096 Stolen AI API Keys

> SOCRadar reports a staging server exposing its /tmp directory over HTTP port 8099 (1,600+ files, ~700 MB), revealing one operator's compromise of a Turkiye-based cryptocurrency exchange with database-layer fund theft and forged withdrawals, plus an automated pipeline harvesting 44,096 unique AI API keys (Gemini, OpenAI, OpenRouter, Anthropic, xAI). Drain amounts are attempted thefts, not verified losses.

- **Published:** 2026-11-17T12:30:09Z
- **Last reviewed:** 2026-11-17T12:30:09Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-6243
- **Card image:** https://intel.threadlinqs.com/og/threat/TL-2026-6243.png
- **ID:** TL-2026-6243
- **Severity:** HIGH
- **Category:** DATA_BREACH
- **Status:** ACTIVE
- **Detections:** 9 (1 shown free below) · **IOCs:** 25 (3 shown free below)

## Description

On 2026-11-17 SOCRadar published passive analysis of an open directory at 176.65.148.2:8099 that served the host's /tmp directory (1,600+ files, ~700 MB). Systemd service directories, X11 sockets, temporary Chromium profiles and a live tmux session directory indicate an active operator workstation rather than a staged collection. The operator toolkit lives in /root/argon-hunt/ and contains more than 75 attack scripts. SOCRadar assesses a single operator or very small team based on consistent coding style, numbered retry files, a single-worker job queue and one toolkit directory. The staging server had reportedly been up since July 2026.

Campaign I targeted a Turkiye-based cryptocurrency exchange (name withheld). The operator performed subdomain brute forcing, port scanning and fuzzing to map internal API routes (user listing, deposits, SMS, withdrawals). The admin panel stored staff passwords as unsalted MD5; of about 100 hashes obtained, 17 were cracked (adms.pot). Hydra found a valid admin login in about one second (log dated 2026-11-12). SMS OTP codes were brute forced (all 10,000 four-digit codes across 12 threads; a variant targeted the six-digit reset code). The operator queried for high-balance customer accounts whose bcrypt hashes used cost factor 4, and sprayed 21 credential pairs across 13 MySQL hosts, obtaining credentials for more than a dozen MySQL accounts including one with read/write access to production. Public CONNECT proxies were used to reach IP-restricted database ports.

Remote code execution and internal pivoting used two database-layer techniques. rce1.py redirected the MySQL general query log to a file in the web root, wrote PHP via a SELECT statement, then requested the file over HTTPS to confirm execution, trying 12 candidate document roots. ssrf.py created an InfluxDB subscription pointing at an internal URL and wrote a data point so the database issued the HTTP request, reaching internal admin panels; a callback captured on the staging server confirmed it. The operator also stole a copy of the admin panel and CRM source code and the exchange's Git history.

Fund theft bypassed the application: drain scripts connected directly to the production database, selected accounts above a balance threshold, replaced the saved withdrawal address with the operator's wallet, and inserted a withdrawal request for 95 percent of the balance plus a matching withdrawal record with origin address 127.0.0.1, created in a pending state flagged for approval. Supporting tools pushed requests through (admin panel credentials, the SSRF route, and a JavaScript payload calling the withdrawal verification endpoint). Output files list USDT: 15 entries (~253,000 USDT), BTC: 19 entries (~4.7 BTC), and a combined sweep of 169 entries (~55,600 USDT, 15.87 ETH, 0.134 BTC); the largest single forged entry was 76,000 USDT and more than 200 forged records were logged. SOCRadar did not verify on-chain settlement, so these are attempted thefts, not confirmed losses. One script added 2.5 BTC to an account before draining it, possibly balance-forgery testing. Watcher scripts retried withdrawal verification when the victim service recovered.

Campaign II is an automated AI API key harvesting pipeline: a Go regex scanner (tcmod/main.go) scans text for OpenAI, Google and Cohere key formats and generic JWTs, followed by validation and numbered recheck rounds (recheck-*.tsv recording HTTP status, rate limits, usage and free-tier flag). 44,096 unique keys were collected: Gemini 26,633 (~6% alive), OpenAI 14,504 (<2% alive), OpenRouter 1,873 (~41% alive), Anthropic 983, xAI 107, plus 519 Google Cloud tokens with broader scope, including administrative and service account keys. On the latest recheck 2,686 keys (6.1%) returned HTTP 200. Per-provider counts sum to 44,100, slightly above the stated unique total (source does not explain the gap). A log of intercepted API traffic held a live service account token, so scanning is not the only collection method. The destination of the keys is not shown.

The operator also maintained a wider target list: 281 exchange domains and 56 entries across target lists (exchanges, banks, payment providers), including 16 key targets (the confirmed victim and several large international exchanges) with headquarters in UAE, Singapore, Turkiye, Brazil, United States and others. Activity included certificate transparency monitoring for four more Turkish exchanges, domain recon on five more Turkish platforms and full subdomain enumeration for a Brazilian exchange. SOCRadar observed no compromise beyond the primary victim. Attribution: Turkish-language artifacts mostly come from the victim environment, browser automation sets a Turkish locale and Europe/Istanbul time zone (possibly imitation of victim customers), and an unrelated bilingual Persian/English project sits in the same directory. Operator origin is not established. Severity is analyst-assigned; the source gives no CVSS and no CVEs. SOCRadar states victim-specific indicators were shared through disclosure channels and are not published, and no file hashes were listed.

## MITRE ATT&CK

- T1595 Active Scanning
- T1190 Exploit Public-Facing Application
- T1078 Valid Accounts
- T1110.001 Password Guessing
- T1110.002 Password Cracking
- T1110.003 Password Spraying
- T1552.001 Credentials In Files
- T1528 Steal Application Access Token
- T1213 Data from Information Repositories
- T1505.003 Web Shell
- T1090.002 External Proxy
- T1596.003 Digital Certificates
- T1590.002 DNS
- T1059.006 Python
- T1021.004 SSH
- T1565.001 Stored Data Manipulation
- T1657 Financial Theft

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**MySQL General Log Web Shell Write (rce1.py) — Exploitation** (`TL-2026-6243-SPL-001`)

Detects the Operation ArgonHunt (TL-2026-6243) database-layer RCE technique from rce1.py: SET GLOBAL general_log_file pointed at a web-root .php path, followed by a SELECT carrying PHP code so the query log becomes a web shell. Matches MySQL audit telemetry for the log path redirect or PHP payload strings, with a Sigma file-creation variant for mysqld writing .php into a document root.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1505.003, T1190
- **Data sources:** mysql_audit_log, AzureDiagnostics (MySqlAuditLogs), syslog (mysqld), sysmon_for_linux
- **Severity:** critical
- **Confidence:** high

### Splunk SPL

````spl
index=database sourcetype="mysql:audit" earliest=-24h
| eval sql=lower(coalesce(sqltext, sql_text, command, ""))
| eval db_user=coalesce(user, user_name, "unknown"), src=coalesce(src_ip, src, ip, "unknown")
| eval log_path_change=if(match(sql, "set\s+(global\s+)?general_log_file\s*=\s*\S*(var/www|srv/www|nginx|htdocs|public_html|wwwroot|/html)\S*\.(php|phtml|php5)"), 1, 0) ```flags redirect of the MySQL general log into a web root as done by rce1.py in TL-2026-6243```
| eval php_payload=if(match(sql, "\x3c\?(php|=)|eval\s*\x28\s*\$_(post|get|request)|(system|passthru|shell_exec)\s*\x28\s*\$_"), 1, 0)
| eval log_toggle=if(match(sql, "set\s+(global\s+)?general_log\s*=\s*(on|1)"), 1, 0)
| where log_path_change=1 OR php_payload=1
| stats count as events, sum(log_path_change) as log_path_changes, sum(php_payload) as php_payloads, sum(log_toggle) as log_enables, earliest(_time) as first_seen, latest(_time) as last_seen, values(sql) as statements by host, db_user, src
| eval risk=if(log_path_changes>0 AND php_payloads>0, "critical", "high")
| convert ctime(first_seen) ctime(last_seen)
````

### Microsoft KQL

```kql
let logpath_re = @"(?i)set\s+(global\s+)?general_log_file\s*=\s*\S*(var/www|srv/www|nginx|htdocs|public_html|wwwroot|/html)\S*\.(php|phtml|php5)";
let webshell_re = @"(?i)<\?(php|=)|eval\s*\(\s*\$_(post|get|request)|(system|passthru|shell_exec)\s*\(\s*\$_";
union
  (AzureDiagnostics
    | where TimeGenerated > ago(24h)
    | where ResourceProvider == "MICROSOFT.DBFORMYSQL" and Category == "MySqlAuditLogs"
    | project TimeGenerated, Source = "AzureMySQL", Host = Resource, DbUser = tostring(user_s), SrcIp = tostring(ip_s), Statement = tostring(sql_text_s)),
  (Syslog
    | where TimeGenerated > ago(24h)
    | where ProcessName in~ ("mysqld", "mariadbd", "mysql")
    | project TimeGenerated, Source = "Syslog", Host = Computer, DbUser = "", SrcIp = HostIP, Statement = SyslogMessage)
| where Statement matches regex logpath_re or Statement matches regex webshell_re
| summarize Events = count(), FirstSeen = min(TimeGenerated), LastSeen = max(TimeGenerated), Statements = make_set(Statement, 5) by Source, Host, DbUser, SrcIp
```

### Sigma

```yaml
title: MySQL Server Process Writing PHP File Into Web Root (General Log Web Shell)
id: 5b0f3c1e-8a47-4d2b-9c61-2e7a4f90b1d3
status: experimental
description: Detects the mysqld or mariadbd process creating a PHP file inside a web document root, the end result of the MySQL general_log_file redirect plus SELECT-written PHP technique used by rce1.py in Operation ArgonHunt (TL-2026-6243).
references:
  - https://socradar.io/blog/argonhunt-crypto-breach-stolen-ai-api-keys/
  - https://attack.mitre.org/techniques/T1190/
  - https://attack.mitre.org/techniques/T1505/003/
author: Threadlinqs Team
date: 2026-11-17
tags:
  - attack.persistence
  - attack.t1505.003
  - attack.initial-access
  - attack.t1190
logsource:
  product: linux
  category: file_event
detection:
  selection_proc:
    Image|endswith:
      - '/mysqld'
      - '/mariadbd'
  selection_file:
    TargetFilename|contains:
      - '/var/www/'
      - '/srv/www/'
      - '/usr/share/nginx/'
      - '/htdocs/'
      - '/public_html/'
      - '/wwwroot/'
    TargetFilename|endswith:
      - '.php'
      - '.phtml'
      - '.php5'
  condition: all of selection_*
falsepositives:
  - 'Database export tooling that legitimately writes files into a web-served directory (rare)'
level: critical

```

**Known false positives:**

- DBAs temporarily enabling the general log to a non-web path for troubleshooting
- Applications that legitimately store PHP source in database rows being logged by audit

**Tuning notes:** Confirm the field names produced by your MySQL audit plugin (sqltext/sql_text, user, src_ip) and adjust the coalesce lists. Extend the web-root path list to your actual document roots. Where possible also alert on any SET GLOBAL general_log_file from non-DBA accounts and remove SUPER/FILE privileges from application accounts.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- InfluxDB Subscription Creation for SSRF ([indicator]) — Exploitation (locked: unlock with TL Creds)
- Web Authentication and OTP Brute Force / Password Spraying (Hydra, [indicator]) — Exploitation (locked: unlock with TL Creds)
- Exposed Git History / [indicator] Harvesting and ffuf Fuzzing — Exploitation (locked: unlock with TL Creds)
- ArgonHunt Operator Python/Go Script Execution — Post-Exploitation (locked: unlock with TL Creds)
- Hydra, ffuf and massdns Tool Execution — Post-Exploitation / Recon (locked: unlock with TL Creds)
- External Source Connecting to Database Ports (Proxy-Tunnelled MySQL/InfluxDB Access) — C2 / Lateral Movement (locked: unlock with TL Creds)
- Forged Withdrawal Rows Written Directly to Exchange Database — Post-Exploitation (locked: unlock with TL Creds)
- Operation ArgonHunt Infrastructure and Drain Wallet Sightings — C2 / Lateral Movement (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sample indicators (3 of 25)

Defanged, TLP:CLEAR.

- IP address: `176.65.148[.]2`
- Domain: `test[.]phoenixbot[.]lol`
- IP address: `176.65.148[.]228`

Full set of 25 indicators unlocks with TL Creds: https://threadlinqs.com/tl-creds

## Sources

- [Operation ArgonHunt: Open Directory Exposes a Crypto Exchange Breach and 44,000 Stolen AI API Keys (Malware News mirror of SOCRadar report)](https://malware.news/t/operation-argonhunt-open-directory-exposes-a-crypto-exchange-breach-and-44-000-stolen-ai-api-keys/127367)
- [SOCRadar original article: ArgonHunt crypto breach and stolen AI API keys](https://socradar.io/blog/argonhunt-crypto-breach-stolen-ai-api-keys/)
- [SOCRadar AI Exposure Checker](https://socradar.io/free-tools/ai-exposure-checker)
- [MITRE ATT&CK T1505.003 Server Software Component: Web Shell](https://attack.mitre.org/techniques/T1505/003/)
- [MITRE ATT&CK T1565.001 Data Manipulation: Stored Data Manipulation](https://attack.mitre.org/techniques/T1565/001/)
- [MITRE ATT&CK T1657 Financial Theft](https://attack.mitre.org/techniques/T1657/)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-6243
