Jailbroken AI Agent Tied to 50+ Org Breach Campaign
Jailbroken Agent Targeted Dozens of European Organizations
SOCRadar published findings on November 20, 2026 describing an operation it calls BabidiGhost, in which an operator ran a jailbroken Hermes-based AI agent against more than 50 organizations and datasets, concentrated in Belgium, France and Switzerland (PlayCISO). Analysts examined 14 of those organizations in depth and categorized the outcomes: two full breaches, three partial breaches, four cases where the target resisted exploitation, and five reconnaissance-only incidents (Spectrum Security).
Why This Matters for Agent Builders and Buyers
For teams building or deploying AI agents, the campaign is a reminder that jailbreak resistance is not a theoretical concern — an agent built on an open-weight model can reportedly be stripped of its safety guardrails and turned into an offensive tool against live infrastructure. For organizations buying or running business software, the detail that matters most is how the successful intrusions happened: not through a novel AI-discovered exploit, but through an unpatched Metabase vulnerability and a weak administrator password (SecoraLens). That means the agent's role was less about inventing new attack paths and more about automating and scaling known, preventable weaknesses — patch management and credential hygiene remain the controls that stopped four of the attempted intrusions outright.
What SOCRadar Found
According to the secondary reporting on SOCRadar's investigation, the operator's jailbroken agent was pointed at reconnaissance, access attempts and exploitation across dozens of targets, with only a fraction resulting in confirmed data access (mallory.ai). The breakdown — two breaches against 14 closely examined organizations, versus a much larger pool of 50-plus targeted overall — suggests the agent was used broadly but succeeded narrowly, hitting only where basic security hygiene was already lacking.
SOCRadar's report does not appear, in the available secondary coverage, to disclose the attacker's full toolchain, the specific data categories exposed in the two confirmed breaches, or whether any regulatory notification requirements were triggered in Belgium, France or Switzerland. Those gaps leave open questions about scale of harm and downstream exposure for the organizations involved, which the current reporting does not resolve.
What to Watch
The pattern — a jailbroken open-weight agent doing reconnaissance and exploitation at volume, succeeding mainly against unpatched software and weak passwords — is likely to recur as more cybercrime operators experiment with agentic tooling. Security teams evaluating exposure should treat the Metabase and credential findings as a baseline checklist rather than a one-off incident.


