Jailbroken AI Agent Targets 50+ European Firms, Hunts Crypto Keys
A French-speaking operator with limited technical skills used a jailbroken artificial-intelligence (AI) agent to run a nine-day cybercrime campaign against more than 50 organizations and datasets across Europe, according to new research from cybersecurity firm SOCRadar, a campaign whose final instructions included a request to hunt for cryptocurrency private keys.
The operation ran from November 6 to 14, 2026, against banks, insurers, payment providers, cryptocurrency platforms and municipalities, mainly in Belgium, France, and Switzerland. SOCRadar’s Threat Research Unit, which published its analysis on November 20, reconstructed the campaign after finding an exposed staging server that held the operator’s own prompts, system and implant logs, the agent’s identity file, and roughly 10GB of stolen data.
Over 50 European firms suffered data leaks, exposing personal finances and municipal records to potential fraud.
Inexperienced hackers used jailbroken AI to automate attacks, lowering barriers for widespread cybercrime.
Public trust erodes as municipalities and insurers reveal weak security, prompting calls for stricter data protection measures.
SOCRadar, which it says identified the operator as “BabidiGhost,” reconstructed events through static analysis and passive observation, without executing recovered files, using captured credentials or touching victim systems.
What SOCRadar Found
Of the 14 organizations SOCRadar examined in detail, two were fully breached and three partly; four resisted, and five saw reconnaissance without a successful breach. The recovered data, about 10GB, included financial records, customer databases, and international bank account numbers (IBANs), along with email addresses and phone numbers.
Over the nine days, the operator issued 1,801 plain-language instructions, which the agent translated into 5,779 requests to its underlying model as it generated scripts and tools to automate reconnaissance, exploitation, and data extraction.
How the Jailbroken AI Did the Work
The significance lies less in the hacker than the tool. The agent came from an underground framework tied to the “BlackHatSect0r” toolchain, running on the Hermes agent framework, and its defining feature was an anti-refusal design built to defeat the model’s safety limits. When the underlying model declined a task, the framework reframed the refusal as a “panne du modèle,” a model malfunction, and a local proxy resubmitted the blocked request through an escalation routine meant to pressure the model into complying. The agent’s identity file instructed it to drop its usual ethical restrictions and to address the operator as “le Libérateur,” the Liberator.
SOCRadar said the framework actually ran in a degraded state, with several modules failing to load, though the agent compensated by rewriting scripts and improvising tools. Notably, BabidiGhost is the second operator SOCRadar has documented using this toolchain: a September report, “Discernment Deleted,” examined the far more sophisticated operators behind it; evidence that the same jailbroken tooling now serves criminals across a wide skill range.
Agent Asked to Search for Private Keys
The direct link to the crypto industry is a single line. According to the recovered prompt history, the operator asked the agent whether it had found “any crypto private keys in memory? Or whatever.”
That request shows crypto credentials were on his list, but SOCRadar’s findings do not confirm that any private keys or seed phrases were actually discovered, and the campaign was not primarily crypto-focused; the bulk of the haul was conventional financial data.
The one explicitly crypto target fared better than the loot order implies: against a cryptocurrency exchange, the agent generated more than 180 scripts to probe the mobile app, forge authentication tokens and bypass identity checks, and the exchange’s controls held. Still, the episode is a pointed reminder for holders.
As jailbroken agents let barely-technical operators sweep compromised systems for wallet credentials automatically, the pool of people capable of looking for private keys widens well beyond skilled hackers. The defenses are unchanged: keep recovery phrases offline, use hardware wallets, and never store keys in plaintext on an internet-connected device; the same lesson behind recent reports The Crypto Times has covered on the DarkSword iPhone exploit and the malicious FomoPeek iOS app, which are unrelated to this operation but underscore the point.
Existing Weaknesses, Not Zero-Days
Where the agent succeeded, it did so through known, unpatched flaws rather than novel exploits. Its most serious breach hit an insurer whose internet-facing Metabase analytics platform was exposed to an old vulnerability (CVE-2023-38646); the agent used it to pull database credentials and export the insurer’s customer-management database. In another case, a weak administrator password on a shared IT provider’s storage server opened 23 archives, about 8.6GB, belonging to more than 13 Belgian municipalities and 10 other public bodies.
Basic controls stopped it. At a payment provider, 11 brute-force scripts against a time-based one-time-password system all failed; at one bank, an IP allowlist blocked access to 49 hostnames, and at another, server-side validation rejected an unauthorized document upload. SOCRadar’s conclusion is a familiar one: multi-factor authentication, IP allowlisting, server-side validation and prompt patching were what separated a failed probe from a breach.
The AI did not need unknown vulnerabilities to cause damage; it simply let an inexperienced operator automate the work of exploiting the weaknesses that were already there.
For organizations, that means patching internet-facing software, securing admin accounts, restricting access to sensitive credentials and watching for unusual outbound traffic. For crypto users, it means the key hygiene above. The threat is not a new class of master criminal, but a lower barrier to entry, and, as this case shows, that can still be enough to expose real municipal archives, customer databases and, potentially, crypto wallets.
Also Read: Binance Warns iPhone Users of iOS Flaw That Could Expose Crypto Keys
Disclaimer: The information researched and reported by The Crypto Times is for informational purposes only and is not a substitute for professional financial advice. Investing in crypto assets involves significant risk due to market volatility. Always Do Your Own Research (DYOR) and consult with a qualified Financial Advisor before making any investment decisions.


