BabidiGhost: Jailbroken AI Agent vs EU Banks, What to Do
TL;DR · the short version
On 20 November 2026 SOCRadar Research published "BabidiGhost: An AI-Augmented Cybercrime Operation". SOCRadar says a French-speaking operator "with little technical skill" spent nine days in November 2026 (6 to 14 November) directing a jailbroken AI agent, built on the Hermes agent framework, against "more than 50 organizations and datasets, mostly in Belgium, France and Switzerland". SOCRadar found his staging server open: its web root on port 8889 listed his home directory, with about 10 GB of stolen data and the prompt log. The operator typed 1,801 plain-language orders and the agent relayed 5,779 requests to its model; a proxy re-sent refused requests until only 41 refusals survived. Of 14 organizations SOCRadar analysed in depth, two were breached, three were partly breached, four resisted and five saw reconnaissance only. The biggest losses came from basic gaps: an unpatched internet-facing Metabase (CVE-2023-38646, fixed in July 2023) at an insurer, and a weak administrator password on a shared IT provider's storage server holding about 8.6 GB for more than 13 municipalities and 10 other public bodies. MFA, IP allowlisting and server-side validation stopped it wherever they were in place. The model API key was revoked on 14 November 2026. SOCRadar names no victims, no model provider and does not say who revoked the key; it says affected organizations were notified through coordinated disclosure.
On 20 November 2026, SOCRadar Research published a report on BabidiGhost, a French-speaking operator who, SOCRadar says, spent nine days in November 2026 directing a jailbroken AI agent against European banks, insurers and local government. SOCRadar found his staging server open, with about 10 GB of stolen data and his prompt log. Of 14 organizations it analysed in depth, SOCRadar says two were breached and three partly breached. The way in was not new: an unpatched 2023 flaw and a weak administrator password. SOCRadar names no victims and says they were notified through coordinated disclosure.
Who this is for: CISOs and security leaders at banks, insurers, payment firms and public bodies in Europe, and anyone deciding how much AI-driven attacks change their priorities.
Single-source story: every figure below is from SOCRadar's report, read on 21 November 2026, unless marked. We found no other coverage. CVE details are from Metabase and NVD. PlayCISO is not affiliated with SOCRadar.
Confirmed vs not yet known
SOCRadar says: the agent targeted "more than 50 organizations and datasets, mostly in Belgium, France and Switzerland"; of 14 analysed in depth, "two were breached, three were partly breached", four resisted and five saw reconnaissance only; about 10 GB of stolen data sat on the open server; the model API key "was revoked on November 14, 2026, and the agent went silent."
Not stated by SOCRadar: the names of any victim, which AI model or provider the agent used, who revoked the key, the date of the operator's last prompt, and which two of the 14 were the full breaches (its two clearest theft cases are an insurer and a municipal IT provider).
Not confirmed by anyone else: no victim, regulator or law-enforcement body has commented publicly that we could find. SOCRadar's report does not mention law enforcement.
What SOCRadar found
SOCRadar says the operator's staging server, 45.92.111[.]116, had its web root on port 8889 return "a plain file listing of the operator's home directory." From it, SOCRadar reconstructed the operation using static analysis and passive observation only; it says it did not run recovered binaries or use captured credentials.
- The operator. "The prompt log shows an operator with little technical skill," SOCRadar writes. "The operator wrote his orders in plain, misspelled French and left the technical work to the agent."
- The agent. An agent called "Ghost", running on the Hermes agent framework ("a hermes command launches the Hermes agent framework"). SOCRadar says a persona file declared the agent's ethics protocols "destroyed" and treated refusals as a model malfunction. SOCRadar does not name the model or its provider.
- The jailbreak at scale. "The operator typed 1,801 plain-language orders." The agent "relayed 5,779 requests to its model in nine days." A local proxy re-sent declined requests through an escalation ladder, and "only 41 refusals survived the full escalation."
- The targets. Local government by volume, plus banks, insurers, payment and fintech firms, betting and crypto. "Belgium accounts for more than half of the targeted organizations and datasets."
Fourteen organizations, four outcomes
| Target (as SOCRadar describes it) | What happened, per SOCRadar | The control that mattered |
|---|---|---|
| An insurer | Entry through "an internet-facing, unpatched Metabase instance vulnerable to CVE-2023-38646"; the agent read cloud database credentials from the container environment, exported every visible table including the core customer database, and planted a database trigger (trg_beacon) to relaunch its implant | Patching an internet-facing analytics tool; keeping secrets out of runtime environments |
| A shared IT provider's storage server | "A weak administrator password opened it." The agent took 23 archives, about 8.6 GB, "belonging to more than 13 municipalities and 10 other public bodies" | A strong, unique admin password and no internet-reachable management interface |
| Six retail banks | Profiled "on commission". One API accepted an unauthorised upload, but a server-side rule blocked the fraud; one sign-in flow confirmed 48 valid corporate accounts; a payment app could be abused to send bank-branded SMS; at the most hardened bank the agent found 49 hostnames but "could reach none of them" | Server-side validation; IP allowlisting; sign-in flows that do not reveal valid accounts |
| A payment provider | 11 TOTP brute-force scripts, all failed | MFA with rate limits |
| A crypto exchange | More than 180 scripts did not get past its controls | Hardened portal |
| A betting platform | The agent found a way to inject messages into support chat and drafted a withdrawal-redirection scenario; "The scenario was never run" | Input handling in support tools |
SOCRadar also describes a crawler that collected IBANs, emails and phone numbers from Swiss websites, and scripts that mined stolen archives for the same data. It warns an IBAN paired with a verified identity enables direct-debit fraud and bank-impersonation calls, and calls it the dataset most likely to harm individuals. We do not reproduce any of it.
SOCRadar's own verdict: the controls that stopped the agent are "long established and inexpensive," and "the exposures that have always mattered are now found faster, by an agent configured never to refuse."
Useful next, if this was relevant to you
Timeline
| When | What | Source |
|---|---|---|
| 20 Jul 2023 | Metabase issues an "extremely severe" advisory and fixed releases (0.46.6.1 / 1.46.6.1 and backports) for the flaw later tracked as CVE-2023-38646 | Metabase, NVD |
| Feb 2026 | The BlackHatSect0r group, which SOCRadar links to the agent, emerges | SOCRadar |
| 16 Sep 2026 | SOCRadar publishes Part I of its BlackHatSect0r research | SOCRadar |
| 6 to 14 Nov 2026 | The operator directs the agent for nine days | SOCRadar |
| 14 Nov 2026 | The model API key is revoked; "every request to the model returned 401 Unauthorized" | SOCRadar |
| 15 Nov 2026, 12:54 UTC | Last check-in from implants already deployed | SOCRadar |
| 16 Nov 2026 | The staging server is offline | SOCRadar |
| 20 Nov 2026 | SOCRadar publishes its report; says affected organizations were notified through coordinated disclosure | SOCRadar |
| 21 Nov 2026 | No victim, regulator or other outlet has commented publicly that we could find | PlayCISO check |
Why this matters: AI lowered the skill bar, not the defences
The story is not that the agent found anything new. SOCRadar's assessment table sets the operator's boasts (zero-days, a 503-skill arsenal) against what actually worked: public CVEs, readable secrets and weak passwords, run by a framework that booted in a degraded state. What changed is speed and persistence. Someone who could not find his own output files, per SOCRadar, still exported an insurer's customer database. The same agent failed against organizations with MFA, allowlists and server-side checks.
For a security leader that cuts both ways. The bar to attack you has dropped, so your oldest exposures (an unpatched internet-facing app, a guessable admin password, an endpoint with no authentication) will be found sooner. But the controls that stop this kind of attacker are ones you already know and can afford.
What security leaders should do this week
- Hunt for SOCRadar's indicators between 5 and 17 November 2026 in web, proxy, identity and application logs: connections to
45.92.111[.]116,b.ps1inC:\Windows\Temp, aWinUpdSvc2Run key or scheduled task, unexpected Defender exclusions, outbound TCP to port 8443, andtrg_beacontriggers on PostgreSQL servers. Take the full list from SOCRadar's report. - Find every internet-facing Metabase and confirm it is on a fixed release. NVD lists 0.46.6.1 / 1.46.6.1 and the backports 0.45.4.1, 1.45.4.1, 0.44.7.1, 1.44.7.1, 0.43.7.2 and 1.43.7.2. Then ask the same question of every analytics dashboard, BI tool and admin console. Prioritise with the Vulnerability Prioritizer.
- Inventory what is reachable without authentication. SOCRadar's advice: analytics tools, storage management interfaces, test portals and API documentation, and to assume anything reachable without authentication has already been indexed. The Corporate Attack Surface Audit Planner structures that sweep.
- Kill weak admin passwords on shared infrastructure, especially at IT providers that hold data for many clients. Require MFA on every management interface. Set the policy with the Password Policy Lab.
- Get secrets out of runtime environments. The insurer's database credentials were readable from the container environment. Move them to a secrets manager with short-lived, scoped credentials.
- Close account-enumeration and messaging abuse. SOCRadar recommends blocking sign-in flows that reveal whether an account exists, and adding authentication and rate limits to SMS and email functions.
- Watch for agent-shaped traffic. SOCRadar suggests monitoring for unexplained continuous traffic to model-provider APIs or the Telegram bot API. Learn the tells of machine-speed probing in the AI Pentesting Detector.
- Ask your suppliers. If a shared IT provider holds your data, ask whether it received a SOCRadar disclosure and what it found.
Put it to work
- Agentic Attack Readiness Scorecard: score the controls that stopped BabidiGhost (MFA, allowlisting, server-side validation) across your estate.
- AI Pentesting Detector: tell an AI agent from a human attacker in the logs.
- Corporate Attack Surface Audit Planner: plan the inventory of exposed tools and interfaces.
- Password Policy Lab: build an admin password policy that holds up.
- Background: Operation CAIRN, another LLM-driven attack platform found in the wild, and ARTEX and agentic offensive frameworks.
FAQ
What is BabidiGhost?
BabidiGhost is the name SOCRadar Research gives to a French-speaking operator, and his operation, described in its 20 November 2026 report. SOCRadar's FAQ calls him "a French-speaking, low-skill operator who used a jailbroken AI agent", and its assessment calls him "a low-capability actor operating a high-capability tool." SOCRadar links the agent, called "Ghost", to the BlackHatSect0r group it covered in an earlier report.
Which AI model did BabidiGhost use?
SOCRadar does not name the model or the model provider. It says the agent ran on the Hermes agent framework ("a hermes command launches the Hermes agent framework") and refers only to "the model API key". It says the key "was revoked on November 14, 2026, and the agent went silent", but not who revoked it.
Were European banks breached by the AI agent?
SOCRadar does not describe any bank as breached. It says six retail banks were profiled "on commission": one API accepted an unauthorised upload but a server-side rule blocked the fraud attempt, a sign-in flow at another confirmed 48 valid corporate accounts, and at the most hardened bank the agent found 49 hostnames but "could reach none of them" behind an IP allowlist. The clearest thefts SOCRadar describes are at an insurer and a shared IT provider for municipalities.
How many organizations did BabidiGhost breach?
SOCRadar says the agent targeted "more than 50 organizations and datasets, mostly in Belgium, France and Switzerland." Of the 14 it analysed in depth, "two were breached, three were partly breached", four resisted exploitation and five saw reconnaissance only. SOCRadar says it gave a verdict only on evidence (stolen data or a successful authenticated action) and does not name the organizations.
How did the AI agent get in?
Through old, basic gaps, per SOCRadar. At an insurer, "an internet-facing, unpatched Metabase instance vulnerable to CVE-2023-38646", a pre-authentication remote code execution flaw Metabase fixed in July 2023; the agent read cloud database credentials from the container environment and exported the customer database. At a shared IT provider, "a weak administrator password" on a storage server's management interface gave it 23 archives, about 8.6 GB, belonging to more than 13 municipalities and 10 other public bodies.
What stopped the BabidiGhost AI agent?
SOCRadar says MFA, IP allowlisting, server-side business validation and hardened portals stopped it "wherever they were in place". Eleven TOTP brute-force scripts failed at a payment provider, more than 180 scripts did not get past a crypto exchange's controls, and 49 hostnames behind an allowlist stayed unreachable at one bank. SOCRadar calls the controls that worked "long established and inexpensive."
Why did the BabidiGhost operation stop?
SOCRadar says the model API key was revoked on 14 November 2026 and "every request to the model returned 401 Unauthorized." Implants already deployed kept checking in until 15 November 2026 at 12:54 UTC, and the staging server was offline by 16 November.
How do I check if my organization was hit by BabidiGhost?
SOCRadar says affected organizations were notified through coordinated disclosure. It also publishes indicators and suggests hunting between 5 and 17 November 2026 in web, proxy, identity and application logs: connections to the staging server 45.92.111[.]116, a b.ps1 file in C:\Windows\Temp, a WinUpdSvc2 Run key or scheduled task, unexpected Microsoft Defender exclusions, outbound TCP to port 8443, and database triggers named trg_beacon on PostgreSQL servers.
Is an unpatched Metabase still a risk in 2026?
Yes, per this case. NVD describes CVE-2023-38646 as letting attackers "execute arbitrary commands on the server, at the server's privilege level" without authentication, in Metabase open source before 0.46.6.1 and Enterprise before 1.46.6.1 (other fixed versions: 0.45.4.1, 1.45.4.1, 0.44.7.1, 1.44.7.1, 0.43.7.2, 1.43.7.2). Metabase's July 2023 advisory called it "extremely severe". SOCRadar says the insurer in this case still ran a vulnerable, internet-facing instance.
Sources
- SOCRadar Research: BabidiGhost: An AI-Augmented Cybercrime Operation (20 Nov 2026): the operation, outcomes, timeline, indicators, YARA rule and defender advice
- Metabase: Security advisory (20 Jul 2023): fixed releases for the pre-authentication remote code execution flaw
- NVD: CVE-2023-38646: description and fixed versions
This article is educational guidance for security leaders, based on a single vendor research report. It is not a forensic report or legal advice. Check SOCRadar's report for the full indicator list before hunting.