Today’s digest is about the deadline you can’t extend: a BI tool that let attackers become admin with no login at all, two unrelated network appliances hitting the same federal patch deadline on the same day, and a state-sponsored group upgrading its tradecraft while defenders are still patching last month’s bugs. Here’s what changed since yesterday’s digest.
Actively exploited watch: A CVSS 10 Metabase SQL injection zero-day gave attackers admin access with no login required
An unauthenticated SQL injection zero-day in Metabase, the popular open-source business intelligence tool, has been exploited in the wild to compromise customer instances and steal data, with Framework and Tally among the confirmed victims. The flaw affects self-hosted Metabase 1.58 and above and lets a remote attacker reach full administrator access without ever authenticating — from there, an attacker can rewrite the application’s database connection settings, harvest the credentials Metabase stores for every connected database, and export whatever data those connections expose. The attack leaves a distinctive two-step signature in access logs: a POST to /api/session/reset_password returning a 400, immediately followed by a GET to /api/user/current returning 200. Metabase Cloud instances were patched automatically; safe minimum versions for self-hosted deployments are 0.58.24, 0.59.21, 0.60.17, 0.61.11, 0.62.9, and 0.63.5. The Hacker News · BleepingComputer
If you self-host Metabase, patch to the safe minimum version for your branch today and treat any database credentials it held as compromised — rotate them rather than assume the attacker didn’t get that far. Then grep your Metabase access logs for the reset-password/current-user pattern above to check whether you were already hit before you patched.
Vulnerability watch: Two unrelated critical flaws hit the same CISA federal patch deadline today
Two critical, unauthenticated command-injection vulnerabilities in widely deployed network appliances both carry an August 10 remediation deadline for U.S. federal agencies under Binding Operational Directive 26-04 — the same day, from two unconnected vendors. CVE-2026-8037 (CVSS 9.6) in Progress Kemp LoadMaster lets an unauthenticated attacker run arbitrary commands through unsanitized input on multiple management endpoints; CISA added it to the Known Exploited Vulnerabilities catalog on August 7 after telemetry logged 792 exploitation attempts from 65 unique IPs across 18 countries over the prior 41 days. CVE-2026-16812 (CVSS 10.0) in Arista VeloCloud Orchestrator on-prem lets an attacker with nothing more than network reach to the web interface execute commands and establish persistence across the entire SD-WAN fleet the orchestrator manages; it was added to KEV in late July. Neither requires a password — both require only a route to the management interface. The Hacker News · BleepingComputer
Neither of these is new — both have had patches available for weeks — so if you run Kemp LoadMaster or VeloCloud Orchestrator on-prem and haven’t patched, today’s deadline is a signal that CISA considers the exploitation risk high enough to force the issue, not a reason to wait for a customer-specific push. Confirm neither management interface is reachable from the public internet regardless of patch status.
AI security watch: North Korea’s Kimsuky group is building local LLM infrastructure to keep its AI use off outside servers
Researchers at Genians found that the North Korea-linked Kimsuky operation has assembled infrastructure for running AI models entirely locally — including Ollama, GPT4All, and Msty, paired with retrieval-augmented generation — alongside AI agent development frameworks, speech-to-text tools, and the AI-assisted coding tool Cursor. Running models locally instead of calling out to hosted AI services means Kimsuky’s operators can feed sensitive stolen or reconnaissance data into an LLM without that data ever leaving infrastructure they control, sidestepping the usage logging and abuse detection that hosted AI providers apply. Genians also recovered finance- and cryptocurrency-themed decoy documents that appear to be AI-generated, consistent with the group’s established phishing playbook. The findings suggest Kimsuky is moving past AI-written lures toward integrating local models into malware development, data analysis, and attack automation directly. Report — Genians
The defensive takeaway isn’t a patch — it’s a detection gap. Threat intel and DLP tooling built around flagging traffic to known hosted AI APIs won’t see anything when the model runs on attacker-controlled infrastructure; if Kimsuky is a threat actor in your sector’s profile, this is worth a note that “no traffic to known AI providers” no longer means “no AI-assisted tradecraft.”
Also noted: Levi Strauss discloses a vishing breach that stole corporate data from three employee machines
Levi Strauss & Co. disclosed in an SEC Form 8-K filed August 7 that attackers used social engineering — industry reporting points to vishing, voice calls impersonating IT or help-desk staff — to gain access to three employees’ computers and exfiltrate corporate data. The company says its response contained and terminated the unauthorized access quickly, that no consumer data was affected, and that it does not expect a material impact on the business; a third-party investigation is ongoing. No technical details of the social-engineering pretext have been released. BleepingComputer · The Record
Three compromised machines from a phone call, not an exploit, is the reminder that your help desk’s identity-verification process is a security control worth testing on the same cadence as your externally facing systems — a callback-to-a-known-number policy before any credential reset or remote-access grant closes most of this attack path.
Final thought
Today’s stories split cleanly into two categories: flaws that were entirely preventable with a patch already sitting in the queue, and a breach that no patch would have stopped because it targeted a person instead of a server. Both need the same discipline — tracking what’s outstanding and closing it before someone else finds it first, whether that’s a CVE or a gap in your identity-verification process. If you want help finding which of your internet-facing management interfaces or help-desk procedures carry that kind of exposure, see our how-to guides or book a discovery call.