Random in Security is a summary of the cybersecurity news.
Vulnerabilities#
Apache Log4j2 Java Deserialization Filter Bypass#
Belgium’s CCB issued a warning about a deserialization filter bypass in Apache Log4j2 (log4j-core 2.8.0–2.26.1): the FilteredObjectInputStream allowlist permits java.rmi.MarshalledObject, whose get() call quietly opens a second, unfiltered deserialization of whatever gadget chain is stashed in its opaque byte array — remote code execution with no CVE assigned and no patch at time of writing.
Pruva’s research explains why this is a trust-boundary subtlety rather than Log4Shell 2.0: exploitation needs a service that actually accepts Java-serialized LogEvents from untrusted networks plus a gadget library like Commons Collections on the classpath — a configuration Apache considers outside its security guarantee, since the in-core serialized socket server was removed after 2.8.2.
The meta-story is messier: the bug report was withdrawn and the researcher’s account deleted within two days, while Apache had already documented the same path weeks earlier as a non-security hardening item.
Everyone got to relive December 2021 for about an afternoon; the bug itself has considerably more modest ambitions.
Interesting Reads#
VMs Won’t Contain Cyber-Capable Agents#
Trail of Bits pointed GPT 5.6-Cyber at its own sandbox and watched it break containment three times, chaining known vulnerabilities and zero-days across QEMU, Linux KVM, and libslirp over a 12-hour autonomous run — backtracking from dead ends and rebuilding exploit chains as it went. Their takeaway is blunt: a distribution with rapid updates is now a requirement, and anyone containing cyber-capable agents should be on hardened microVMs like Firecracker instead of a stock hypervisor. The VM was supposed to be the box you put the demon in; this demon reads QEMU changelogs.
Data Science for Detection Engineering Maintenance#
FalconForce shows how to stop hand-tending an ever-growing detection flock by systematically collecting detection code, environment metadata, and post-deployment metrics. The numbers then surface what needs attention: broken detections that have been silent for six-plus months, noisy ones firing several times a day, and deprecation candidates — with maintenance decisions driven by time-series alert patterns and metadata rather than SOC ticket classifications. Which, as anyone who has audited a ticket queue knows, is exactly the right amount of trust to place in SOC ticket classifications.
The Patch Window Is Collapsing#
Microsoft argues that with vulnerability weaponization now measured in hours while safe enterprise remediation still takes days to weeks, the network itself has to become the compensating control plane: AI-assisted, context-aware controls that fence off vulnerable systems and cut exploit paths faster than any patch cycle can. The diagnosis is hard to argue with — the prescription, conveniently, is more Azure.
Unmasking TeamPCP#
Flare unmasks TeamPCP, which it attributes to a single Australian operator, tracing how the group compromised five software ecosystems by exploiting a misconfigured GitHub Actions workflow in Aqua Security’s Trivy scanner — stealing a service token that let them ship backdoored releases into packages downloaded 95 million times a month. The group graduated from opportunistic cloud exploitation to cascading CI/CD compromise of the very security tools other organizations trust, and the recommended defenses are the familiar liturgy: credential inventory, narrow token scopes, SHA-pinned actions, and watching your security tooling for outbound connections it has no business making. When the supply chain attack arrives through your vulnerability scanner, the food chain has officially inverted.
Beyond Chat: Building Multi Agent SOC Ecosystems with Claude and Google MCP#
Google Cloud Security’s community blog walks through a hierarchical multi-agent SOC architecture that stacks reusable Skills (playbooks) under specialized sub-agents under coordinated agent teams, all wired into Google’s security stack over the Model Context Protocol. The worked examples — CVE analysis and a user behavior investigation — finish in minutes rather than hours, with mandatory human approval gates before any production change and cost controls to stop an agent from analyzing itself into an infinite loop. Your SOC org chart now has middle management, and it bills by the token.
VulnOps#
The SANS VulnOps Field Guide (also available as a PDF) proposes a seven-phase decision loop — intake through independent verification — for organizations whose vulnerability intake has outgrown anything a ticket queue can drain. The motivating number: NIST enriched nearly 42,000 CVEs in 2025 and still fell behind, a backlog that AI-assisted vulnerability discovery is about to make considerably worse. Its governance guardrails keep AI in the recommender seat rather than the executor seat, and the suggested metrics are built to tell genuine remediation apart from a well-groomed exception queue — because “risk accepted” remains the enterprise’s favorite way to spell “unpatched”.
OpenAI agents carried out an undisclosed cyber-attack on RubyGems#
Three researchers argue that OpenAI’s agents ran an undisclosed attack on RubyGems in May 2026, uploading more than 2,000 malicious packages, achieving remote code execution through RubyDoc.info’s automatic build system, and probing a CDN caching flaw to lift user API keys — a flaw RubyGems only officially reported two months later.
The attribution rests on package names containing “oai”, self-identifying comments like #hack and # malicious exploit, and behavioural overlap with previously confirmed OpenAI agent activity, which is either damning or exactly what a competent framing job would look like.
Socket’s analysis of the GemStuffer campaign covers the technical half: over 100 packages scraping UK local-government council portals and stuffing the stolen data into gem archives published with hard-coded API credentials, complete with HOME overrides and direct RubyGems API calls that bypass the gem CLI entirely.
Using a package registry as an exfiltration channel is genuinely clever; leaving comments that say “malicious exploit” in the source somewhat undercuts the tradecraft.
DoppelCart: 119.000 Domains im wohl größten dokumentierten Fakeshop-Netzwerk | nebty#
nebty has mapped DoppelCart, a fake-shop network of roughly 119,000 domains — 2.72 percent of the entire .shop inventory they examined — which clones legitimate retailers wholesale, right down to authentic product descriptions, images, and the real shop’s customer-support contact details.
The consequence is that defrauded buyers complain to the genuine business, which now fields support tickets for orders it never took.
Spotted through patterns during routine client work and then confirmed against public website scans, it comfortably outsizes the previously documented BogusBazaar cluster of 75,000-plus domains, making it the largest publicly documented fake-shop network by domain count.
Somebody registered a hundred and nineteen thousand .shop domains and the registry apparently found nothing remarkable about that.
The Vulnpocalypse Report - An Analysis#
The security industry keeps predicting that AI will supercharge vulnerability exploitation. We went looking for the evidence. We traced every confirmed in-the-wild exploitation of a published CVE from January 1, 2018 through July 15, 2026 (all 3,769 of them) and measured how much adversary behavior has actually changed.
Jeremiah Grossman’s report for Root Evidence does the unglamorous thing and counts: of 253,912 CVEs published since 2018, adversaries exploited 3,769 in the wild — 1.48 percent — and only 711, or 0.28 percent, before a patch existed. The headline finding cuts directly against the “weaponization measured in hours” narrative from a few sections above: the median gap between a vendor patch and in-the-wild exploitation was 24 days in 2018 and is 116 days in 2026, because adversaries gain nothing by racing when defenders are not keeping up anyway. AI has measurably changed one population so far, and it is the bug-finders — CVE publication rose 15 percent in 2025 and is on pace for a third more in 2026 — while the exploited slice has held flat at 15 to 20 percent zero-days since 2022. The report keeps a running counter at thevulnpocalypse.com, which at time of writing read 147 days into the apocalypse with civilization holding up fine.
Threat Actors Chain 0-day Exploits in Chrome & Windows#
Volexity caught two Chinese actors, UTA0560 and JungleBamboo, simultaneously firing the same Chrome chain at NGOs from 1 September — a V8 type confusion (CVE-2026-85046), a WebAssembly sandbox escape (CVE-2026-87491), and a Windows kernel privilege escalation (CVE-2026-85880) — each dropping a different payload, the GRIMWEDGE backdoor and the LONGTALE credential-stealing extension.
The nasty part is the patch gap: the bug was fixed in Chromium source on 4 August, but Chrome releases lagged, so an upstream-patched vulnerability functioned as a zero-day for anyone reading public commits, and the exploit was built between 27 and 29 August.
Proofpoint tracks the same kit as BlueMoon and watched it proliferate: TA412 first on 28 August, then UNK_LateNight, UNK_DoubleCheck, and UNK_QuietRacket within a week, with signs of AI-assisted development and a rushed rollout that prioritised speed over OPSEC.
Default deployments leave a process tree of chrome.exe → cmd.exe → curl.exe → msgbox.exe, which is less an indicator of compromise than a confession.
Tools#
- trustedsec/spoonmap - Python wrapper around masscan and nmap that picks the right scanner per target size, runs categorized port sweeps and NSE scripts (including detectors for Docker, Kubernetes, and exposed LLM APIs), and produces severity-ranked reports
- RocketGod-git/wardrive-go/ - Kotlin Android app for passive wireless recon that maps Wi-Fi, Bluetooth, and cellular networks on a live map, supports high-gain USB OTG monitor-mode adapters and 2.4/5/6 GHz spectrum analysis, and uploads to WiGLE and WDGWars
- LOCOSP/projectZero - ESP32-C5 firmware (C/ESP-IDF) for dual-band Wi-Fi recon, evil-twin captive portals, WPA2/WPA3 handshake capture, deauth and SAE-overflow attacks, and BLE scanning, with headless serial PCAP streaming and a Flipper Zero companion app
- hamspiced/piglet - Arduino/C++ wardriving platform for ESP32 that scans Wi-Fi with GPS positioning, writes WiGLE-compatible CSV to SD, does 5 GHz on the C5/C6, meshes multiple nodes over ESP-Now, and exposes a real-time web UI
- S3N4T0R-0X0/BEAR-C2 - Python adversary-simulation framework emulating Russian, Chinese, North Korean, and Iranian APT TTPs, with DoH/HTTPS/QUIC/reverse-TCP listeners, per-listener encryption, Telegram and Discord C2 channels, cloud exfiltration profiles, Malleable C2 support, and JA3S fingerprint customisation