Russian Intelligence Turns IP Cameras and Routers Into a NATO Surveillance Grid

Dutch intelligence and allied agencies revealed a large-scale Russian espionage operation abusing internet-exposed cameras and routers to watch military shipments to Ukraine, and the EU and UK answered with their first joint cyber sanctions against the GRU. Beyond the state-actor news, CISA flagged two Joomla zero-days under active exploitation and researchers demonstrated fresh prompt-injection attacks that weaponize binaries and images against AI coding and reverse-engineering agents.

Threat Activity

• Russian state hackers are exploiting insecure internet-connected IP cameras and doorbell devices near NATO military bases and transport routes to monitor Western arms deliveries to Ukraine, according to the Dutch AIVD and MIVD services. Cameras pointed at logistics corridors in the Netherlands, Ukraine and other NATO countries were accessed to identify what equipment was being sent to Kyiv (Rijksoverheid, Yahoo/AIVD). A parallel NCSC-UK advisory warns critical-sector operators of the same actors’ worldwide targeting of poorly configured routers (NCSC-UK).
• The EU and UK issued their first joint cyber sanctions package, hitting dozens of Russian individuals and entities and accusing the GRU of coordinating a network of hacking groups behind attacks across Europe (BleepingComputer).
• A misconfigured phishing server exposed three active Evilginx operations targeting Microsoft 365. An operator left python3 -m http.server 8080 running with directory listing on, and French firm Lexfo lifted the entire toolkit from a readable .bash_history, pivoting to two additional campaigns (The Hacker News).
• New ransomware crew “D1R” surfaced listing Synopsys, Arm and Bosch on its leak site. Trackers note the group appears to be leveraging a single supply-chain incident at a major semiconductor design-software provider to name multiple downstream victims (FalconFeeds, DarkFeed).
• DeadLock ransomware is running at a high operational tempo, publishing 10+ new victims in a matter of days including Indonesian plastics maker Vinilon Group (FalconFeeds, DarkFeed).
• Lazarus reportedly used CVE-2024-21338 as a zero-day with no BYOVD required — the vulnerable driver already ships on every Windows machine, removing the need to drop one (@cr3ghost via Dinosn).
• APT37 continues its cloud-based C2 model with RokRAT variants abusing pCloud, Dropbox and Yandex, reusing the same Yandex OAuth account and TTPs seen in prior Operation Artemis activity (blackorbird).
• A LightSpy iOS malware sample tied to Chinese state activity surfaced on VirusTotal, extending the mobile-surveillance toolset’s tracking (@Now_on_VT via cyb3rops).
• Armenian national Karen Vardanyan pleaded guilty to aiding Ryuk ransomware attacks that extorted over $15M from US organizations between 2019–2020 (SecurityAffairs).

Vulnerabilities & Exploits

• CISA added two maximum-severity (CVSS 10.0) Joomla extension flaws to KEV following zero-day exploitation. CVE-2026-48939 (iCagenda) and CVE-2026-56291 (Balbooa Forms) both allow unrestricted file upload leading to web shells; admins should update iCagenda to 4.0.8/3.9.15 and Balbooa Forms to 2.4.1 and hunt for rogue PHP files and admin accounts (The Hacker News, CISA).
• A pre-auth RCE path in Cisco Unified CM (CVE-2026-20230) turns the phone system into a foothold. The WebDialer SSRF allows unauthenticated code execution via crafted SOAP requests and cluster-wide persistence; the researcher notes current patches only add URI filtering while other weaknesses remain (SecureLayer7).
• Unauthenticated RCE disclosed in Motorola’s MR2600 router, with a full write-up detailing the chain (mrbruh.com).
• Dell BIOS passwords can be recovered from SPI flash (CVE-2026-40639) due to reversible XOR “encryption,” enabling fast recovery of short passwords and partial recovery of longer ones — full BIOS control on physical access (AmberWolf).
• A Realtek driver flaw lets user-mode code abuse the DMA controller without additional hardware or a second driver (Dinosn).
• The Windows GDID (Global Device Identifier) has been fully reverse-engineered, and Microsoft acknowledges Windows 11 ships the tracker with no off switch — the same identifier first surfaced publicly in an FBI complaint used to correlate a suspect’s device (earlier coverage). Full technical write-up via blackorbird.

Offensive Techniques & Detection

• COM hijacking has become the dominant lateral-movement primitive of the last few years, per iPurple, who points to yet another COM hijack targeting Chrome and Edge. @Cyb3rMonk notes DLLs loaded into legitimate processes via COM are far more evasive to EDR than launching EXEs (iPurple).
• Malicious AMSI providers can be registered via regsvr32 to execute code during PowerShell input. iPurple’s playbook covers detection through registry, process-creation and DLL-load telemetry, and recommends enabling AMSI signing enforcement as mitigation (ipurple.team).

AI & Model Security

• US Navy researchers turned binaries into prompt-injection weapons against AI reverse-engineering agents. Malicious prompt strings embedded as ordinary C string variables get fed into LLM-backed tools like Cline and GhidraMCP during decompilation, making the agents misreport what a program does while the binary runs unchanged — indirect prompt injection pushed down to the binary level (@0x0SojalSec).
• Someone is actively scanning for exposed MCP servers and AI-assistant credentials, per the SANS Internet Storm Center — a reminder that agentic infrastructure is now an enumerated attack surface (SANS ISC).
• OpenAI is offering $50,000 for a universal jailbreak of ChatGPT’s biosafety challenges, formalizing red-team incentives around its safety guardrails (@interesting_aIl).
• Frontier models remain prone to hallucinated and injected outputs on adversarial images. Researchers showed GPT-5.6 Sol and Claude Fable 5 confidently “reading” nonexistent hidden messages and meaningless scribbles; notably, a 2023 viral “picture of a rose” prompt injection appears baked into Fable’s weights as the canonical image-injection response (@goodside, @fabianstelzer).
• An AI chatbot platform, EcoGPT, allegedly exposed shared conversations via an unauthenticated Supabase shared_conversations table, with a claimed 13,255 users’ chat histories dumped to JSON/CSV — unverified (DailyDarkWeb).

New Tools & Releases

• LIEF 1.0.0 shipped, the binary-parsing library adds expanded format support, a new runtime API, API-safety features and C++23 components — a staple for reverse engineers and tooling authors (lief.re).
• Colibri is a zero-dependency mini inference engine written in pure C that streams MoE experts from disk, letting consumer PCs with ~25 GB RAM run the 744B-parameter GLM-5.2 without a GPU or Python runtime (blackorbird).

Data Breaches

• IKEA Norway disclosed a breach exposing over 45,000 customer and employee records. Attackers accessed roughly 5,000 IKEA accounts in a GitHub development system via a third-party vulnerability, downloading 3,261 employees’ data and 42,370 customers’ names and addresses; the true scope may be larger (IS.fi).
• Dutch nationals are suspected in the Odido mobile breach that exposed data on some six million customers (via Dinosn).
• Centers Laboratory breach affects 540,000 individuals, with extortion group WorldLeaks claiming to have stolen 720 GB from the healthcare-testing provider (SecurityWeek).

AI Industry & Policy

• Anthropic extended free Claude Fable 5 access for paid subscribers through July 19, delaying the switch to pay-per-use — widely read as a response to pricing pressure from OpenAI’s GPT-5.6 Sol (The Decoder).
• OpenAI temporarily relaxed GPT-5.6 Sol usage limits after demand surged over 48 hours (BleepingComputer).
• A German consortium released Soofi S, an open 30B-A3B hybrid MoE trained entirely on Deutsche Telekom’s Munich cloud that tops fully-open competitors on both German and English benchmarks (The Decoder).
• Google’s SensorFM foundation model was trained on over a trillion minutes of wearable data from five million Fitbit and Pixel Watch users, beating benchmarks on 34 of 35 health tasks (The Decoder).
• Dutch oversight body CTIVD found intelligence agencies trained AI on citizen data they weren’t authorized to hold, raising fresh questions about data governance in state AI programs (@T3chFalcon via Dinosn).