A Patch That Missed, a DNA File You Can Edit, and a List With Your Organisation's Name On It

On 24 March 2026, two malicious LiteLLM releases, v1.82.7 and v1.82.8, went live on PyPI and stole credentials from anyone who installed them. Five months later, CloudSEK published a searchable lookup built from roughly 434,000 captured files, mapping potential exposure to more than 2,500 named organisations.

The malicious code in v1.82.8 ran at Python interpreter startup, so it fired whether or not you ever used LiteLLM. In addition, the exposure is now public and searchable by name or domain, which means your customers and regulators can look you up before you do.

If you build AI features on an LLM gateway, check your March install window and rotate anything that was reachable.

I have written this one up in full, including the indicators of compromise and the remediation steps.

1. A Thermo Fisher flaw lets an attacker edit DNA test results without detection

CVE-2026-17583, CVSS v4.0 8.2. An attacker can modify the .fsa and .hid output files from Applied Biosystems genetic analyzers, and the files themselves will not show it. Thermo Fisher has added digital signatures for five product lines. Three end-of-life products get no fix and need compensating controls instead.

Read the alert here

2. CISA added four vulnerabilities to the KEV Catalog in early August, one because a patch failed

N-able N-central is the one to act on first. CISA added CVE-2026-18556 and CVE-2026-18577 to the KEV Catalog in early August, both letting an attacker reach an admin account without credentials. Build 2026.3.1.10 closes it. If your provider says they patched N-central and confirm.

The other two, IBM Langflow (CVE-2026-9198) and Apache Tomcat (CVE-2026-34486), give attackers unauthenticated paths to compromise and both are being exploited now.

Read the alert here

3. CISA, the FBI and HHS updated the Medusa ransomware advisory, and named healthcare as a frequent victim

Updated 18 August with new tactics and indicators from FBI investigations through April 2026. HHS joined as a co-sealer. More than 500 organisations hit so far. The recommended controls are network segmentation and traffic filtering. If you have been deferring a segmentation review, here is the business case.

Read the alert here

4. August vendor patch cycles: Microsoft, SAP and Oracle all shipped, and Oracle changed its schedule

Microsoft fixed 421 CVEs on 11 August, 62 of them Critical, across Windows, Office, SharePoint, Exchange and Azure. SAP shipped 28 new notes the same day, including CVE-2026-58231 at CVSS 10.0 in Commerce Cloud and two critical code injection flaws in Manufacturing Integration and Intelligence, which runs on regulated plant floors.

Oracle's Critical Security Patch Update from 18 August covers 943 patches across 23 product families, more than 460 of them exploitable without credentials. Two are worth prioritising: CVE-2026-70862 (CVSS 9.1) in Application Testing Suite, reachable over plain HTTP without authentication, and CVE-2026-70715 (CVSS 8.8) in Autonomous Health Framework's Trace File Analyzer, reachable over TLS from adjacent networks. No product in this batch is a clinical system by name, but Oracle infrastructure commonly sits underneath regulated environments, so check whether any of the 23 affected families run in yours before assuming this one skips you.

Microsoft | SAP | Oracle's advisory | Full Oracle write-up

1. NIST wants to hear how you are using AI in your CSF work, and comments close 15 October

SP 1353 is out as an initial public draft, a QuickStart Guide on applying AI tools to CSF 2.0 analysis and reporting. No MedTech tie in the document. The reason to care is that how NIST frames AI assisted control mapping now will shape what an auditor accepts as evidence later.

Read the announcement here

2. FDA wants feedback on how to regulate generative AI in medical devices, and comments close 19 October

On 18 August the FDA opened a comment period on a discussion paper covering risk assessment, premarket evaluation, and postmarket monitoring for generative AI enabled medical devices. If you build or plan to build generative AI into a regulated device, this is the chance to shape the framework before it hardens into guidance.

BEFORE YOU CLOSE THIS

Four things worth doing this week.

1. Check whether LiteLLM was installed anywhere in your stack on 24 March, and rotate what was reachable.

2. Get the N-central build number from whoever runs your RMM. It needs to be 2026.3.1.10 or later.

3. Confirm who owns CRA incident reporting in your organisation before 11 September.

4. Check whether any of Oracle's 23 affected product families run in your environment, and if so, prioritise CVE-2026-70862 and CVE-2026-70715 first.

What struck me putting this together is how ordinary each gap was. A patch that missed. A file nobody could verify. A compromise that sat quiet for five months. None of it needed a sophisticated attacker, just something nobody had gotten around to checking.

Thanks for reading the first edition.

My goal here is simply to bring you the news that actually affects your day to day. If one item above did that, it worked.

See you in the next one.
Andrew.