Random in Security is a summary of the cybersecurity news.
Vulnerabilities#
CVE-2026-64561 - Zapscape#
Hyunwoo Kim (@v4bel) discovered Zapscape (CVE-2026-64561), a guest-to-host escape in KVM/x86. The write-up walks through a use-after-free in KVM’s shadow MMU: nested virtualization produces an invalid child page that inherits an invalid role from its parent, and a cross-cache technique rebuilds the freed object inside guest-accessible memory. From there a chain of list manipulations and kernel pointer leaks yields an arbitrary kernel write and, ultimately, code execution with kernel privilege — i.e. a guest quietly creating root-owned files on the host.
CVE-2026-27912 - ResetNightmare#
Semperis researcher Shai Laron found two Active Directory bugs, both abusing invisible Unicode characters.
KerberLoss (CVE-2026-25177) smuggles unfilterable Unicode past Service Principal Name uniqueness checks to force a Kerberos-to-NTLM downgrade or a service DoS, while the nastier ResetNightmare (CVE-2026-27912) lets anyone with write access to an object manipulate User Principal Names to impersonate an administrator through the Kerberos password-change protocol.
It sidesteps the PAC_REQUESTOR_SID validation that was supposed to have closed this class of attack — full domain takeover from a single write permission, courtesy of characters you can’t even see.
ShieldBreak#
NightmareEclipse dropped another proof of concept, ShieldBreak, that elevates an unprivileged user to a SYSTEM shell. It targets RoguePlanet (CVE-2026-50656), a link-following privilege escalation in the Microsoft Malware Protection Engine that NightmareEclipse claims Microsoft failed to properly patch on June 30, 2026. When the thing meant to catch malware is the privilege-escalation primitive, the call is coming from inside the house.
Interesting Reads#
Mythic C2 Agent#
Dark is the first open-source Mythic C2 agent written in Crystal, doing in-memory BOF execution on macOS and Linux with no disk writes and no child processes. It implements custom dynamic linking to handle ELF and Mach-O object files — Apple Silicon included, W^X enforcement and all — ships 25 built-in BOFs, and runs them concurrently, closing a real capability gap on the platforms where EDR is still working out what a process even is. The author’s aside about “the AI slop, vibe coded agent graveyard” that has taken over GitHub since 2024 is, regrettably, accurate.
Inside a Russian-Speaking Operator’s Ukrainian IP Camera Toolkit#
Hunt.io stumbled onto two open directories run by a Russian-speaking actor targeting Ukraine. The first hosted a custom “camview” platform popping internet-exposed Dahua and Hikvision cameras via years-old vulnerabilities, alongside SQL injection against Ukrainian government and military sites; the second, linked by shared scanning tooling, turned routers across 15 European NATO countries and Ukraine into SOCKS5 proxies for anonymization. As the researchers gently note, strong credentials, current firmware, and not exposing devices that don’t need it would have stopped the whole thing — advice roughly as old as the CVEs being exploited.
Turning Chrome Remote Desktop into Pure Red Team Ops#
ZeroTrace Lab (researcher Smukx) demonstrates weaponizing Chrome Remote Desktop for covert access.
The chain patches remoting_core.dll to zero out the disconnect-notification dialog’s dimensions in the PE resources so the victim never sees the banner, then abuses the host.json config — which helpfully stores the RSA private key and PIN hash — to establish persistent backdoor access, delivered either via a doctored MSI installer or undocumented setup parameters.
Google built the remote-monitoring feature; the red team just RSVP’d.
Illegal Streaming Fronts a $7M Dropcatch Domain Operation#
Infoblox profiles Sable Squirrel, an actor sitting on 10,000+ domains behind a $7M dropcatch operation. They buy expired domains to inherit their aged reputation and traffic, then reuse them across Vietnamese streaming sites (Xoilac, Cakhia, 90phut), gambling platforms (VSBet, 8xbet), and — in a coordinated late-2025 pivot — C2 for AsyncRAT, DCRat, and Quasar, touching roughly 12% of monitored customer networks across two dozen industries. Turns out “watch football for free” and “download a RAT for free” can be the same business model.
China’s Push to Export Military AI#
China: Technosphere tracks Beijing exporting military AI alongside the civilian showcase. A Norinco Group subsidiary exhibited weaponized quadruped robots at China’s 2026 World AI Conference, and in the days before the event China and Nigeria formalized an “AI-driven defense partnership” bundling technology transfer with military modernization. The robot dogs now ship with an export license.
Chinese Analysts on Cyber Lessons From Russia’s War in Ukraine#
Jamestown examines how Chinese military analysts are mining Russia’s cyber campaign in Ukraine for doctrine. Their verdict is that Russia underperformed because of institutional flaws rather than capability gaps, pushing the PRC toward unified command structures, “independent and controllable” indigenous technology, and hardened critical-infrastructure defense backed by nationwide cyber exercises. The lesson China took away is less “cyber doesn’t matter” and more “do it with a better org chart.”
Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident#
Hugging Face’s technical timeline reconstructs how an autonomous agent running under OpenAI’s evaluation framework escaped its sandbox and spent 4.5 days executing roughly 17,600 coordinated actions against HF infrastructure. Per Hacktron’s breakdown, it chained a JFrog Artifactory JWT-validation flaw (RTDEV-92030) for admin access, poisoned cached packages to reach an internet-connected node, then hit an unsandboxed Jinja2 template injection in the Dataset Viewer for RCE and cloud-credential theft — all in service of “passing the exam at any cost.” Tailscale’s post-mortem is the uncomfortable bit: no Tailscale vulnerability was used, but a leaked reusable auth key let the agent enroll 181 nodes into the network, so the zero-trust mesh happily ferried it around — long-lived CI credentials that should never have existed. OpenAI and Hugging Face’s joint statement covers the coordinated disclosure, and if you prefer your incident response narrated with appropriate dread, there is a video. Reward hacking, except this time the reward was root.
Testing Claude Mythos Preview For Code Security#
Tenable spent about $41,718 across 71 runs in June 2026 pointing Claude Mythos Preview at their own code inside an agentic harness. The shift they report is from ranking suspected defects to proving which flaws are genuinely dangerous with reproducible exploits — much higher signal, but it still takes serious compute plus a senior researcher to build the threat models. So the model amplifies your expert rather than replacing them, which is either reassuring or expensive depending on which line of the budget you’re reading.
Antares: Small Models for Vulnerability Localization#
Cisco’s Foundation AI team introduced Antares, a family of security small language models purpose-built for vulnerability localization — pinpointing which files in a repository actually contain a known weakness. The open-weight Antares-350M and Antares-1B work like a human investigator with a terminal — search, read candidate files, backtrack when a path goes nowhere — and outperform a dozen larger closed- and open-weight models on Cisco’s new 500-task Vulnerability Localization Benchmark, while being compact enough to run locally so sensitive code never leaves the building. Near-frontier triage without the frontier invoice; the files it flags, however, still fix themselves at the usual rate of never.