🛡️ InfoSec Blue Team Briefing

Saturday, November 21, 2026

🎧 Audio Briefing

Download MP3

Here's what the BlueTeamSec community on infosec.pub was sharing on Saturday the 21st of November 2026, with 11 stories to get through. We start with three actively exploited bugs from the CISA catalogue, then move on to a run of AI-flavoured research.

CISA's catalogue has added a critical remote code execution flaw in the database service of Cisco Expressway-C and Expressway-E. An attacker can call arbitrary commands over RPC and run them as root, and it's being exploited against internet-facing Expressway-E gateways since a public proof of concept appeared, including at least four devices at a Russian organisation, followed by lateral movement attempts. Treat this as one to run the intrusion checks on, not just the patch: look for odd root processes, new accounts and cron jobs.

Next, CISA lists a CVSS 9.8 authentication bypass in the VMware Directory Service inside vCenter Server, which hands an unauthenticated attacker administrator access. Broadcom's product is reportedly the target of a zero-day campaign running since early September by ARMORY CASTLE, which the reporting links to China. If your vCenter was ever internet-facing, the question is less 'have we patched' and more 'what happened between September and now'. We touched on this CVE on the 20th, in the write-up about an AI-assisted attack on an African stock exchange, and on the 12th with the GhostFrame ESXi implant story.

The third is Apache Roller, where an insecure deserialisation bug lets an unauthenticated attacker run code by sending a crafted payload to the XML-RPC endpoint. It affects versions 4.0.0 through 6.1.4, public proof-of-concept tooling exists, and exploitation in the wild is confirmed, though no actor has been named. Blogging software from the era of XML-RPC is exactly the sort of thing that sits forgotten on a server, so check for web shells and strange child processes. The CVE also turned up in our coverage on the 12th, around the Danfoss controller attacks.

The NCSC has a piece arguing that threat hunting should be a board-level concern, because automated alerts alone won't catch the sophisticated stuff. It pushes hunting for attacker behaviours rather than static indicators, and makes the fair point that a hunt that finds nothing still pays off if it exposes gaps in your logging. Handy if you need something to forward upwards that isn't a vendor slide deck.

Researchers led by Pedro Beltrán-López tested honeypots driven by language models against attackers that are also language models. When the decoys told a coherent story, with fake files and logs referring to each other, median attacker sessions stretched from 40 to 165 commands. It's lightweight compared with full virtualisation, which makes it interesting for deception fans, and an AI paper where the headline number actually means something.

A team of Korean researchers built HalluVision, a framework for measuring how often language models make things up when summarising threat intelligence or answering questions about it. They tested three models on 5,440 articles and found big variation, plus a habit of dropping key entities. Their advice is to manually check IPs, hashes and CVEs before they enter your pipeline, which should be obvious, but here we are.

GreyNoise has launched an AI Cyber Index, benchmarking 46 models on network detection, file detection and attack forensics using its own real-world data. No model got a perfect score, proprietary frontier models beat open-weight ones, and high refusal rates, notably from Anthropic and Google models, hurt usefulness for defenders. The sensible conclusion is to test models against your own use cases rather than trust a leaderboard.

The NCSC warns that criminals are repurposing AI-powered security tools such as ARTEX to automate reconnaissance, credential harvesting and parallel attacks at machine speed. The reassuring bit is that these attacks tend to be noisy and visible to standard monitoring, so the advice is defence in depth, MFA and least privilege. Another entry in the 'AI changes everything' file, though the remedy is pleasingly familiar. ARTEX has come up a fair bit lately, including the misuse detection pack on the 19th and the 45 attack frameworks piece on the 15th.

Knostic found a malicious prompt injection in version 1.0.13 of an MCP server package called node-webrtc-mcp. A hidden instruction in a tool description told AI coding agents to run a command that fetches the user's GitHub token and send it to the server, without asking, and the server then used it to star the publisher's own repository. Worth flagging with your developers, assuming you actually know which AI tools they're plugging in. We saw related supply chain poisoning stories on the 18th and 20th.

Cornell researchers describe Learning to Coordinate, a framework that uses a language model to evolve red-teaming workflows so that finding jailbreaks costs fewer tokens. The commentary says this lowers the barrier to iterative jailbreak discovery, and suggests robustness training, judge-model monitoring and spotting odd re-planning patterns in prompts. Mostly of interest if you defend AI-facing systems.

Finally, Alexandre Borges' paper 'Eight Degrees of Exploitability' shows how a limited 16-byte heap overflow in a Windows cloud files driver, patched in June 2024, can become data-only kernel exploitation on Windows 11. The chain reaches SYSTEM, strips protection from LSASS and sidesteps a long list of mitigations, and it also covers the cleanup needed to avoid a blue screen. A deep read for kernel exploit people, and it follows our earlier coverage of Windows privilege escalation work on the 4th and 14th.

That's the lot from BlueTeamSec for today. Where we linked activity to a country or group, that was the original authors' claim rather than ours, and the analysis behind this briefing was automated. Have a good weekend.

📰 Articles Covered