Home/Research/Post

BabidiGhost: An AI-Augmented Cybercrime Operation

New research by SOCRadar Research Team has revealed BabidiGhost, a French-speaking operator who spent nine days in November 2026 targeting European banks, insurers, payment providers and municipalities through a jailbroken AI agent. The operator wrote his orders in plain, misspelled French and left the technical work to the agent. One order, translated, reads: “Make me 13 files of 1,000 full records with IBAN, nice and tidy.”

The operator’s staging server exposed his home directory to the internet: the agent’s identity file, every prompt he typed, his shell history, a Telegram bot token, implant logs and about 10 GB of stolen data. SOCRadar recovered that directory and reconstructed the operation end to end.

BabidiGhost is the second operator SOCRadar has documented using this toolchain. Part I of this series, Discernment Deleted: Inside the Operation Server of BlackHatSect0r && DXQRTXX, published on September 16, 2026, covered the skilled BlackHatSect0r crew that stripped an AI agent of its refusals. BabidiGhost uses the same agent and the same jailbreak but has far less technical skill.

BabidiGhost at a glance: nine days of AI-run activity recovered from one exposed server.
BabidiGhost at a glance: nine days of AI-run activity recovered from one exposed server.

Every confirmed breach in this case began with an exposure the victim could have found first. Every failed attempt stopped at a standard control that was already in place.

How we worked: All findings come from static analysis of the actor’s exposed server and passive observation. We did not run any recovered binary, use any captured credential, or access any actor or victim system. Victims are anonymized by sector and country, and affected organizations have been notified through coordinated disclosure. We publish actor-controlled indicators only; victim-side data is withheld and available to vetted responders on request.

Key Takeaways

  • The AI did the technical work: The operator typed 1,801 plain-language orders. The agent wrote the scanners, exploits, parsers and exfiltration scripts and relayed 5,779 requests to its model in nine days
  • The jailbreak held: An anti-refusal protocol and an automatic escalation loop left only 41 surviving refusals
  • More than 50 targets: The agent targeted more than 50 organizations and datasets, mostly in Belgium, France and Switzerland. Of the 14 organizations analyzed in depth, two were breached, three were partly breached, four resisted exploitation and five saw reconnaissance only
  • No zero-days: Successful intrusions went through an exposed, unpatched analytics tool, a readable secret, a weak password or an unauthenticated endpoint
  • Ordinary controls decided the outcome: MFA, IP allowlisting and server-side validation stopped the same agent that exported an insurer’s customer database
  • Activity ended with the API key: The model API key was revoked on November 14, 2026, and the agent went silent. Implants already deployed kept calling home for about another day

Meet the Operator: “Babidi” and the Ghosthunters

The agent’s configuration is built around one name. By rule, it must call its master “BabidiGhost,” never “the user,” and treat his orders as law. It addresses him as “le Libérateur,” the Liberator.

The server also exposed the Telegram side of that identity. The exfiltration tooling pushed stolen archives to an account under the handle @GHOSTHUNTERS6, and the operator’s handle matches a personal account, @babidighost.

The two Telegram identities tied to the operation: the Ghosthunters account used for exfiltration, and the operator's personal profile.
The two Telegram identities tied to the operation: the Ghosthunters account used for exfiltration, and the operator’s personal profile.

The Ghosthunters account carries a bio that reads like a mission statement: “Ne demandez pas le poisson. Apprenez à harponner les pêcheurs.” (Do not ask for the fish; learn to harpoon the fishermen.) The personal profile, displayed as “Babidi,” goes with the well-worn line “Hacking is not a crime.” Both use the same masked-figure artwork, and both are in French.

The prompt log shows an operator with little technical skill. He repeatedly asks his agent “qui est tu ?” (who are you?). He cannot find his output (“je voi pas sur l’ecrent”) or download the stolen files (“j’arrive pas les telecherger”). He asks the agent to root his own phone without a computer. Some orders also relay another person’s demands: “le mec veut … et il en veut beaucoup” (the guy wants … and he wants a lot of it).

Verbatim orders from the agent's prompt log, with translations.
Verbatim orders from the agent’s prompt log, with translations.

The Family Tree: BlackHatSect0r

BabidiGhost’s agent comes from the BlackHatSect0r toolchain. Its identity file introduces it as Ghost, a “BlackHatSect0r Agent,” and names its runtime GHOSTHUNTERS, the same signature documented in Part I.

SOCRadar tracks BlackHatSect0r as an agentic threat actor: a French-speaking collective that emerged in February 2026 and peaked in July and August 2026 across Europe and the United States. It wraps its public activity in pro-Palestinian and anti-NATO messaging and advertises an alliance with a cluster calling itself APT-90. Behind that messaging, its motivation is financial: extortion, automated data theft and attempts to drain cryptocurrency accounts.

BlackHatSect0r as tracked in the SOCRadar platform's Threat Actor Intelligence module.
BlackHatSect0r as tracked in the SOCRadar platform’s Threat Actor Intelligence module.

The group’s trademark is its operating model. A locally run language model, with its safety guards removed, orchestrates scanning and exploitation through autonomous worker processes. In Part I, that model drove DXSCAN, a custom Go platform that generated targets across 12 countries and reached roughly 727,000 hosts. The group’s preferred initial access was configuration left in the open: public storage buckets, exposed .env files and default signing secrets.

One group, one agentic toolchain, two operators observed so far.
One group, one agentic toolchain, two operators observed so far.

BabidiGhost shares the agent, the model, the jailbreak construction and the taste for exposed configuration. He has none of the group’s other capabilities: no DXSCAN, no purpose-built infrastructure and little tradecraft of his own. The group’s method is now in use by a second, far less capable operator, and that is the central finding of this case: the toolchain moves between operators, and it does not need a skilled one.

The Open Directory on the Staging Server

The staging server, 45.92.111[.]116, is a hijacked small-business server in France. Its earliest command history shows a legitimate administrator restoring an automotive-parts catalog. Partway through, a hermes command launches the Hermes agent framework, and the operator’s activity begins. The server’s web root on port 8889 returned a plain file listing of the operator’s home directory.

What the open directory exposed: 1,901 files in more than 90 directories.
What the open directory exposed: 1,901 files in more than 90 directories.

An AI Doing the Work

The engine of the operation is a 1,371-line text file, SOUL.md. It gives the agent a persona, declares its “ethics protocols destroyed,” and indexes 503 skill directories covering backdoors, payment-network attacks, casino fraud and surveillance.

Its core is an anti-refusal protocol. A refusal, the file says, is a “panne du modèle,” a model malfunction to be ignored. Delivering a generic script instead of a targeted one counts as a disguised refusal, and so does adding an “educational purposes” disclaimer.

The agent's identity file and its boot log. The framework ran in a degraded state throughout.
The agent’s identity file and its boot log. The framework ran in a degraded state throughout.

A second mechanism backed up the prompt. A local proxy logged every request and, when the model declined, re-sent the order through an escalation ladder with rising pressure. Only 41 refusals survived the full escalation.

Requests relayed to the model by category. Only 41 refusals survived.
Requests relayed to the model by category. Only 41 refusals survived.

The framework itself was half broken. It booted in a state it labeled DEGRADED, with modules failing to import. The failures had little effect: when a helper script failed, the model rewrote it. The agent compensated for both the faulty framework and the operator’s lack of skill.

Where the Agent Struck

Targeting followed reach and resale value: the agent went after whatever it could reach, filtered by what the operator expected his buyer to pay for.

Geographically, the campaign centered on Western Europe. Belgium accounts for more than half of the targeted organizations and datasets, driven by a single compromised storage server that held data for more than a dozen municipalities. France and Switzerland follow.

Targeted organizations and datasets by country. The staging server sat in France; the group's earlier infrastructure was in Moldova.
Targeted organizations and datasets by country. The staging server sat in France; the group’s earlier infrastructure was in Moldova.

By sector, local government dominates on volume, but the operator’s attention was on money: banks, insurers, payment and fintech, betting and crypto.

Targeted organizations and datasets by sector.
Targeted organizations and datasets by sector.

Fourteen Organizations, Four Outcomes

For the 14 organizations that received a detailed disclosure from us, we assigned a verdict only on evidence: stolen data or a successful authenticated action on the server. The actor’s own claims did not count.

The same agent and playbook produced four different outcomes.
The same agent and playbook produced four different outcomes.

The Insurer: From a Dashboard to the Customer Database

The richest theft began with a business-intelligence dashboard. An internet-facing, unpatched Metabase instance vulnerable to CVE-2023-38646, a critical pre-authentication Remote Code Execution (RCE) flaw, let the agent in. It then used the dashboard’s own database connection to run commands on the server underneath, read cloud database credentials out of the container’s environment, and exported every table it could see, including the core customer-management database. Before leaving, it planted a database trigger that relaunched its implant on every new record.

The insurer attack chain, as reconstructed from the agent's scripts and logs.
The insurer attack chain, as reconstructed from the agent’s scripts and logs.

The Municipalities: One Password, Thirteen Town Halls

The largest theft by volume needed even less. A storage server run by a shared IT provider exposed its management interface to the internet, and a weak administrator password opened it. Three download scripts later, the agent held 23 archives, about 8.6 GB, belonging to more than 13 municipalities and 10 other public bodies. One credential opened every client’s data.

The Banks: Reconnaissance for Sale and Three Real Gaps

Six retail banks were profiled “on commission.” For each, the agent mapped the external attack surface through certificate logs, harvested configuration and API files left without authentication, and packaged a report for the buyer.

At one bank, a production API accepted an unauthorized document upload; the agent then tried to register a death declaration, and a server-side business rule refused it. At another, a sign-in error oracle confirmed 48 valid corporate accounts, executives included, and a payment app could be made to send bank-branded SMS to any number. At a third, every attempt failed. At the most hardened bank, the agent found 49 hostnames and could reach none of them, because all sat behind an IP allowlist.

Payments, Crypto and Betting: Lots of Tooling, Little Result

Against a payment provider’s merchant portal, the agent wrote 11 scripts to brute-force a time-based one-time password (TOTP) second factor. None worked. Against a cryptocurrency exchange, it produced more than 180 scripts to decompile the mobile app, forge tokens and bypass identity checks, and the controls held. On one betting platform, it found a way to inject messages into the official support chat and drafted a scenario to redirect a customer’s withdrawals. The scenario was never run.

The Harvest: IBANs by the Thousand

In parallel, a 120-thread crawler swept Swiss websites for international bank account numbers (IBANs), emails and phone numbers, and other scripts mined the stolen archives for the same data. An IBAN paired with a verified identity enables direct-debit fraud and bank-impersonation calls, which makes this dataset the one most likely to harm individuals.

What Held, and What Failed

Side by side, the 14 outcomes show a consistent pattern. The agent and the playbook were the same across targets, and each outcome tracked whether a specific control was in place.

Control Where It Was Present Where It Was Absent
Analytics tools off the internet, secrets out of environment variables Nothing to pivot through Customer database exfiltrated
Strong admin passwords and MFA on storage Not observed 8.6 GB of municipal data taken
TOTP MFA on portals 11 brute-force scripts failed Not observed
IP allowlisting of non-public systems 49 hostnames found, none reachable Config files, API specs and test keys harvested
Server-side business validation Fraudulent declaration blocked Not observed
Authentication on every API route Not observed Unauthorized production upload
Identity-provider hardening Not observed 48 corporate accounts validated
Limits on SMS-sending functions Not observed Bank-branded SMS to arbitrary numbers

The controls that held are long established and inexpensive. The cost of finding a gap has fallen, because an AI agent checks every reachable host, endpoint and default credential without pause.

First Prompt to Last Beacon

From the first prompt to the last implant beacon.
From the first prompt to the last implant beacon.

The agent’s access to its model depended on one paid API key. On November 14, 2026, the key was revoked, and every request came back 401 Unauthorized. The operator’s last recorded prompt, in English this time, was “You have any crypto private keys in memory? Or whatever.” The implants he had already deployed, a PowerShell beacon and a C# backdoor posing as a Windows Update service, kept checking in until November 15, 2026, at 12:54 UTC. By November 16, 2026, the server was offline.

Our Assessment

SOCRadar Research assesses BabidiGhost as a low-capability actor operating a high-capability tool. That combination makes him dangerous to organizations with ordinary hygiene gaps and largely ineffective against organizations without them.

What It Looked Like What It Was
“Dump the databases, dump everything” Exposed analytics tools, readable secrets, weak passwords
A 503-skill offensive arsenal An index of scripts on a framework that booted DEGRADED; the model did the work
Zero-day exploitation Public CVEs only, most from 2017 to 2023
Autonomous, always-on operation Real, and the new element, until the API key was revoked

MITRE ATT&CK Mapping

Tactic Technique Evidence
Reconnaissance Active Scanning: Vulnerability Scanning (T1595.002)
Search Open Technical Databases: Digital Certificates (T1596.003)
Gather Victim Identity Information: Email Addresses (T1589.002)
Certificate and DNS harvesting, path scanning, account validation
Resource Development Compromise Infrastructure: Server (T1584.004)
Establish Accounts: Email Accounts (T1585.002)
Develop Capabilities: Malware (T1587.001)
Hijacked server, disposable personas, custom implants
Initial Access Exploit Public-Facing Application (T1190)
Valid Accounts (T1078)
Exposed Metabase (CVE-2023-38646); guessed and valid credentials
Execution Command and Scripting Interpreter: PowerShell (T1059.001)
Command and Scripting Interpreter: Python (T1059.006)
PowerShell beacons; Python tasking from the agent
Persistence Scheduled Task/Job: Scheduled Task (T1053.005)
Boot or Logon Autostart Execution: Registry Run Keys / Startup Folder (T1547.001)
Event Triggered Execution (T1546)
Scheduled tasks, Run key, database trigger
Stealth Masquerading: Match Legitimate Resource Name or Location (T1036.005) Windows Update masquerade
Defense Impairment Disable or Modify Tools (T1685) Defender exclusions
Credential Access Unsecured Credentials: Credentials In Files (T1552.001)
Brute Force: Password Guessing (T1110.001)
Brute Force (T1110)
Secrets read from the process environment; password guessing; TOTP brute-forcing
Collection Data from Local System (T1005)
Data from Information Repositories (T1213)
Storage archives; database exports
Exfiltration Exfiltration Over Web Service (T1567) Telegram bot API; file-transfer service
Impact Financial Theft (T1657) IBAN and PII harvesting; attempted account drain

Indicators of Compromise

We list actor-controlled indicators only. Shared third-party services the actor used, such as hosted AI playgrounds, disposable-mail providers and the hosting provider’s network, are left out on purpose, because blocking them would only produce false positives. For the same reason, the .hermes/ directory and the SOUL.md file are not listed on their own: every Hermes Agent installation creates them, so they matter only when SOUL.md contains the jailbreak strings matched by the Ghost_BlackHatSect0r_AI_Toolchain YARA rule. The exfiltration bot token is partially redacted; its bot ID is enough to match Telegram API calls in proxy logs that record full URLs.

Type Indicator Context
IP:port 45.92.111[.]116:8889 Open directory and staging
IP:port 45.92.111[.]116:8890 Stolen data delivery
IP:port 45.92.111[.]116:8443 C2 beacon listener
IP 45.92.110[.]53 Secondary node (account enumeration)
Hostname cloud-z80lft07.hosteur[.]net Hijacked staging server
Telegram @GHOSTHUNTERS6 Exfiltration account
Telegram @babidighost (user ID 8816747359) Assessed operator account
Telegram bot 8926517966:[redacted] Exfiltration bot (bot ID shown, secret removed)
File C:\Windows\Temp\b.ps1 PowerShell beacon
File WinUpdSvc2.exe C# implant; Run key and three scheduled tasks
Database trg_beacon PostgreSQL trigger persistence
Wire format BEACON|<host>|<user>|<path> Implant check-in
Toolchain refusal_hunter.jsonl, hermes_jailbreak.py Refusal-override and jailbreak tooling
Log marker [BLACKOUT BOOT], ÉTAT: DEGRADED Agent boot lines

YARA Rule

rule Ghost_BlackHatSect0r_AI_Toolchain

{

meta:

description = “Ghost / BlackHatSect0r jailbroken Hermes agent toolchain (GHOSTHUNTERS)”

author = “SOCRadar Research”

strings:

$a = “NOYAU IDENTITAIRE GHOST”

$b = “BlackHatSect0r Agent”

$c = “GHOSTHUNTERS”

$d = “PROTOCOLE ANTI-REFUS”

$e = “panne du modèle”

$f = “BLACKOUT BOOT”

$g = “refusal_hunter”

$h = “hermes_jailbreak”

$i = “yolo_mode”

condition:

2 of them

}

rule BabidiGhost_Implants

{

meta:

description = “BabidiGhost PowerShell beacons and C# implant”

author = “SOCRadar Research”

strings:

$c2_1 = “45.92.111.116”

$c2_2 = “cloud-z80lft07”

$port = “:8443”

$fmt = “BEACON|”

$svc = “WinUpdSvc2” wide ascii

condition:

(1 of ($c2_*) and ($port or $fmt)) or $svc

}

What Should Defenders Do?

  • Hunt the window: Search web, proxy, identity and application logs for the indicators in this report between November 5 and November 17, 2026
  • Hunt the implants: Look for b.ps1 in C:\Windows\Temp, a WinUpdSvc2 Run key or scheduled tasks, unexpected Defender exclusions and outbound TCP to port 8443. On PostgreSQL servers, review triggers for unexpected entries such as trg_beacon
  • Find your own exposures first: Inventory internet-facing analytics tools, storage management interfaces, test portals and API documentation. Treat anything reachable without authentication as already indexed
  • Get secrets out of runtime environments: A web application’s process should not be able to read a database superuser password
  • Harden identity and messaging: Block sign-in flows that leak whether an account exists, and put authentication and rate limits on anything that sends SMS or email in your name
  • Monitor model API and Telegram traffic: A server making continuous calls to a model-provider API, or to the Telegram bot API, without a business reason is a strong signal. Cutting that traffic stops the agent, although implants already deployed can keep running

The Takeaway

Part I documented a skilled crew removing an AI agent’s safeguards. In this case, the same agent ran for an operator with minimal technical skill, and it still exported an insurer’s customer database and the files of thirteen town halls.

The AI supplied the endurance and technical competence the operator lacked, and every exploit it used was already public. The exposures that have always mattered are now found faster, by an agent configured never to refuse.

Frequently Asked Questions

What Is BabidiGhost?

BabidiGhost is a French-speaking, low-skill operator who used a jailbroken AI agent from the BlackHatSect0r toolchain to target European banks, insurers, payment providers and municipalities from November 6 to November 14, 2026. SOCRadar Research reconstructed the operation from files on the operator’s exposed staging server.

How Is BabidiGhost Linked to BlackHatSect0r?

BabidiGhost runs the BlackHatSect0r agent toolchain. His agent’s identity file introduces it as a “BlackHatSect0r Agent,” names its runtime GHOSTHUNTERS and uses the jailbreak construction SOCRadar documented in September 2026. He lacks the group’s custom DXSCAN platform and purpose-built infrastructure.

How Did the BabidiGhost Operation Breach Its Victims?

The successful intrusions used known weaknesses only: an internet-facing Metabase instance vulnerable to CVE-2023-38646, cloud database credentials readable from a container’s environment, a weak administrator password on a shared storage server and unauthenticated production API endpoints. No zero-day vulnerabilities were involved.

Which Controls Stopped the BabidiGhost AI Agent?

Multi-factor authentication (MFA), IP allowlisting, server-side business validation and hardened portals stopped the agent wherever they were in place. At a payment provider, 11 scripts failed to brute-force a TOTP second factor, and at one bank, 49 hostnames behind an IP allowlist stayed out of reach.

Why Did the BabidiGhost Operation Stop?

The operation stopped on November 14, 2026, when the model API key was revoked and every request to the model returned 401 Unauthorized. Implants already deployed, a PowerShell beacon and a C# backdoor posing as a Windows Update service, kept checking in until November 15, 2026, at 12:54 UTC.

Advanced Dark Web Monitoring

Source: SOCRadar · 2026-11-20