🛡️ InfoSec Blue Team Briefing

Monday, November 16, 2026

🎧 Audio Briefing

Download MP3

BlueTeamSec briefing for Monday the 16th of November 2026, drawn from the community on infosec.pub. We have 14 stories to get through, and a fair few of them have the same autonomous pentest framework lurking in the background, so settle in.

Google Project Zero is scrapping its ninety-plus-thirty disclosure policy for a strictly time-based model. The window drops to sixty days in February 2027 and thirty days in January 2028, whatever the patch status, and it's explicitly a response to AI-assisted bug hunting and patch reverse-engineering. Vendors who treat the patch gap as a leisurely affair have been put on notice, so one for anyone who owns a patch schedule.

Apple has published a short piece on Secure Exclaves, a hardware-enforced architecture in Apple silicon that isolates sensitive workloads from the main kernel. It's a brief precis and the text is truncated, so there's little detail on offer. This sits alongside the Apple threat notifications to officials we covered on the 4th, though it's a different topic entirely. Skim it only if you track Apple platform security.

Cloudflare says autonomous, agent-driven pentest frameworks have exploded since the attempted ARTEX takedown in early October, with over twelve thousand repositories and demand up more than thirty-fold. Thousands of IPs are pointing these tools at its customers, notably abusing stolen CI/CD tokens that never expire. Cloudflare proposes a six-pillar defence model, and the non-expiring tokens are the bit your own house might want to check first.

Barracuda disclosed that on the 1st and 2nd of November an attacker got into an internal cloud environment holding credentials Barracuda XDR uses to integrate with customer services. Entry came via a cloud access key left in a config file in a third-party developer's GitHub repository, and the activity looks consistent with ARTEX, hunting for credentials and crypto wallets. There's no direct evidence customer credentials were accessed, and only Managed XDR Cloud and Monitored Endpoint API integrations were potentially affected. ARTEX has been all over our coverage, including the FBI overview and a Chinese Computer Federation piece on the 15th.

Trezor found that Safe 5 wallets sold via the reseller CryptoBilis had been physically tampered with. A rogue circuit board between the mainboard and screen mirrored the display, capturing recovery seeds and sending them out over a mobile network. Trezor pushed an emergency firmware update, dropped about eighty percent of its resellers and is offering free replacements to anyone who bought from CryptoBilis since March. A tidy reminder that buying hardware wallets from anywhere but the maker is a supply chain decision.

Cursor has written up an incident where an unattributed attacker used an access key leaked by its vendor Fireworks AI to reach a storage bucket, then moved laterally into internal systems. They took archived non-production source code and some support data, and slipped malicious copies of the Claude Code extension into the cache serving the marketplace. Up to five thousand users may have received a bad build on the 29th of October, and the breach was contained by the 1st of November. Another leaked key doing the heavy lifting, which is becoming a theme.

A US criminal complaint charges Zyaire Wilkins and co-conspirators with hiding malware, including cookie and session-token stealers, in video games distributed through a digital platform, reaching thousands of downloaders. The group allegedly used stolen identities to look credible as developers, and investigators matched Web2 service logs with blockchain forensics. The defender takeaways are fairly standard, but the investigative method is the interesting bit. Related reading includes Europol's Operation ENDGAME writeup from the 14th.

Claroc Research found that apps built with xAI's Grok Build feature default to anyone-with-the-link access, leaving 35,578 apps, nearly ninety-five percent of those tested, reachable on grok.me. They belong to xAI-related entities and at least thirty other organisations, including six Fortune 500 firms. Weak server-side checks let anyone read, edit or delete data, and client code carries hardcoded keys. A default of public is a bold choice for something people will put business data in. Railway's incident report on the 14th is loosely adjacent reading.

SOCRadar mapped 916 live ARTEX consoles on US infrastructure as of the 11th of November, around ninety-eight percent over plain HTTP. Two Docker images carry a trojanized login function allowing unauthenticated remote code execution, so people running the attackers' tooling may be getting run themselves. SOCRadar notes the framework has been tied to Storm-0940 supply-chain activity and attacks on South Korean finance. If you wanted evidence that offensive tooling needs its own security review, here it is.

NVISO Security details a tenant isolation bypass in LiteLLM's semantic cache, tracked as CVE-2026-89032 with a CVSS of 8.7. A metadata key mismatch means an authenticated user can send similar queries and pull other tenants' cached responses and tool calls. No active exploitation is reported, and disabling semantic caching is an option. Particularly relevant if you run LiteLLM as a shared gateway.

Intezer shows modern AI agents can rebuild functional source code from compiled binaries and packages, even reconstructing parts of real software like Slack. It's still imperfect and labour-intensive, but it chips away at security through obscurity and licence enforcement. Another entry in the AI-changes-everything file, though this one at least comes with a demonstration. We saw a related agent reverse engineering tool on the 15th.

Suture-pic is an open-source Windows implant framework for offensive engineering and C2 research, not a live threat. It builds position-independent loaders paired with modular payloads, composed and dispatched by name from a fixed registry, with TCP and HTTP comms. It's meant for isolated lab use only. Handy if you follow implant tradecraft, and similar to the Donut fork we covered on the 11th.

Cisco Talos has analysed ARTEX, which uses a planner and worker architecture plus the Model Context Protocol to automate secrets hunting, injection and password spraying. The worry is lower barriers for multi-stage attacks at machine speed against defence, finance, government and critical infrastructure. Recommended defences are behavioural baselining, spotting high-velocity enumeration and automating credential lifecycles. Note that this is a community poster's secondary summary, not the original Talos report.

Thalium has released BinQL, a CodeQL extension that lets you write declarative queries against compiled binaries. It reimplements the extractor and uses decompiler output to abstract away architecture, covering x86, ARM and MIPS. That makes variant hunting in closed-source software, firmware and drivers more scalable. Worth a look if you hunt bugs without source.

And that's your lot for today from BlueTeamSec. Where activity was tied to a particular group or organisation, that was the original authors' claim rather than ours, and the analysis behind this briefing was automated. Check those leaked keys before the weekend, and we'll see you tomorrow.

📰 Articles Covered