45 karma · joined June 16, 2023
By the way, all modern models know which hospital the last pope was admitted to shortly before his death.
I did some intensive testing with changelogs and came to the conclusion that knowledge is partitioned and comes with variable cutoff dates.
Common tools, like GIT, have their feature cutoff in like 2023 or even earlier. Enough for a model to work with the tool when there was no groundbreaking change.
Gossip has a fresher cutoff date. People ask models like an interactive Wikipedia.
Given some similarities, I’d like to briefly mention `krun` here. Although it’s an OCI-compatible container runtime, it uses MicroVMs with a similar approach. Perhaps we can exchange ideas here? I recall that GPU passthrough is also a recurring topic there.
Meh… I think I should work on a PR to fix this…
This summary outlines the key events and remediation actions from the official incident reports published by Localmind.ai between October 5 and October 9, 2025.
Incident overview and initial response (October 5)
On October 5, 2025, at 05:43 CEST, Localmind detected unauthorized access to its systems. The immediate response was to take all affected systems, including internal platforms and customer instances, offline to contain the breach. Initial measures included:
- Resetting all passwords and regenerating API keys (e.g., for Notion, SendGrid, Hetzner).
- Deactivating all user accounts, restricting access to a minimal number of administrators with mandatory two-factor authentication (2FA).
- Initiating a forensic investigation.
The breach originated from a misconfiguration in an externally accessible beta-test instance. The flaw granted administrator privileges by default to a newly registered account. The attacker used this access to:
- Access the integrated automation platform (n8n).
- Retrieve an unrestricted API key for the internal Notion knowledge base, which contained infrastructure documentation and credentials.
- Use the compromised information to escalate access further and send emails from an internal account.
Impact assessment and forensic Updates
- Scope: The core Localmind platform was not compromised. The attack was confined to administrative interfaces and test environments. A limited number of customer systems were accessed, while on-premise instances showed no signs of unauthorized access.
- Forensics: Unauthorized logins were traced to IP addresses from VPN providers, complicating attribution. Login activity occurred outside regular business hours (nights, weekends). As of October 8, no evidence of large-scale data exfiltration was found.
- Data transparency: Localmind offered data exports to customers to conduct their own audits for potential GDPR breach notifications.
The company initiated a comprehensive infrastructure rebuild and security overhaul.
1. New infrastructure: A migration of virtual machines to new, Tier IV, ISO 27001/27018 certified data centers with a fully isolated infrastructure was nearly complete as of October 8. Systems are being rebuilt from clean data volumes (e.g., Docker volumes) onto new, hardened hosts.
2. Access security:
- Implementation of an F5 Web Application Firewall (WAF) with pre-authentication for each customer instance.
- Mandatory two-factor authentication (2FA) for all application logins.
- Deployment of the Wazuh security agent for centralized login monitoring and anomaly detection.
- All previous service accounts and credentials within automation workflows were deleted, requiring a re-issue.
3. Automation restriction: Critical automation nodes in n8n (e.g., Execute Command, Read/Write File to Disk) were disabled and will be unavailable in cloud environments going forward.
4. Enhanced monitoring: Additional security agents were deployed for endpoint security, configuration assessment, file integrity monitoring, and threat intelligence.
5. Process change: Each customer instance undergoes a manual audit and documentation before restart, with the audit protocol provided to the customer.
On October 9, Localmind reported a renewed attempt to gain unauthorized access. The new security measures successfully blocked these attacks. The only confirmed impact was a brief, unauthorized text modification on a separately hosted, external development website, which was promptly reverted. The company attributes this attempt to the same threat actor.
Status as of latest update (October 9, 2025)
Systems were in a phased, controlled restart process, with customers being kept informed. The company continues to work on audits and security fortifications.
Sources (as Mementos)
<https://web.archive.org/web/20250000000000*/https://www.loca...> <https://web.archive.org/web/20250000000000*/https://security...>
About the use case: Ultimately, I just want to enable my team to use different LLMs without having to create an account for each team member. However, it should be possible to upload attachments. RAG is not required at the moment.