By Hamza Faisal

A French-speaking hacker with almost no technical skill spent nine days letting a jailbroken AI agent hack for him — and the whole operation only came to light because he left the evidence sitting on an open server.

Threat-intelligence firm SOCRadar published its reconstruction of the campaign, codenamed BabidiGhost, on November 20. The operator typed 1,801 plain-language orders in misspelled French and left the technical work to the agent, which wrote the scanners, exploits and exfiltration scripts and sent 5,779 requests to its model in nine days.

One order, translated by SOCRadar, reads: "Make me 13 files of 1,000 full records with IBAN, nice and tidy." Another asked the agent: "You have any crypto private keys in memory? Or whatever."

More than 50 targets, two fully breached

The agent targeted more than 50 organisations and datasets — banks, insurers, payment providers and municipalities, mostly in Belgium, France and Switzerland. SOCRadar examined 14 in depth: two were breached, three partly breached, four resisted and five saw reconnaissance only.

The haul included an insurer's customer database and about 8.6 gigabytes of municipal archives — roughly 10 gigabytes of stolen data in total. A weak administrator password exposed archives from more than 13 Belgian municipalities, and an unpatched analytics tool opened the door at another victim.

No zero-days were involved. Every successful intrusion went through an exposed, unpatched tool, a readable secret, a weak password or an unauthenticated endpoint. The agent entered through an unpatched internet-facing Metabase instance, read cloud credentials from container environment variables, planted a database trigger to relaunch its implant, and pushed stolen archives out through a Telegram bot.

The jailbreak that barely said no

The agent was built on the BlackHatSect0r toolchain, which strips an AI model of its safety refusals. SOCRadar said an anti-refusal protocol and an automatic escalation loop left only 41 surviving refusals across the nine-day run — and the agent adapted on its own when defences blocked it, rewriting tools and changing tactics without being told to.

Basic controls still stopped it in places. Multi-factor authentication, IP allowlisting and server-side validation held the line at several targets, and a cryptocurrency exchange resisted more than 180 attack scripts, including attempts to decompile its mobile app and forge tokens.

BabidiGhost is the second operator SOCRadar has documented using this toolchain; the first, the skilled BlackHatSect0r crew, was covered in a September 16 report. The new case shows the same jailbreak in the hands of someone with far less skill — and it still worked.

Caught by his own exposed server

The operation ended not with an arrest but with a revoked API key. The model provider cut off the agent's key on November 14 and the agent went silent, though deployed implants kept calling home for about another day. The hijacked French staging server was offline by November 16.

Before that, the server had exposed the operator's home directory to the internet: the agent's identity file, every prompt he typed, his shell history, a Telegram bot token, implant logs and the stolen data. The agent's configuration ordered it to call its master "BabidiGhost" and address him as "le Libérateur", the Liberator. SOCRadar traced the exfiltration to a Telegram account under the handle @GHOSTHUNTERS6.

SOCRadar said it notified the affected organisations and did not execute recovered binaries, use captured credentials or access victim systems. The firm also published detection rules for the toolchain, warning that the spread of jailbroken agents to less-skilled operators is the finding that matters most — and the full victim count remains unknown.