The tech landscape is reeling from the weaponization of the programmatic ad ecosystem as a delivery pipeline for zero-click mercenary spyware, while engineers push the boundaries of firmware modernization, kernel exploitation on abandoned silicon, and rack-scale emulation.
AdFection: How Real-Time Bidding Became a Weaponized Zero-Click Pipeline
For years, cybersecurity engineers viewed Real-Time Bidding (RTB) exchanges primarily through the lens of privacy degradation—data brokers harvesting device telemetry and location graphs. Today, that narrative changed permanently. A joint investigation by Citizen Lab and Confiant exposed AdFection, a commercial campaign demonstrating the weaponization of the programmatic advertising ecosystem into an automated delivery pipeline for zero-click mercenary spyware.
+------------------+ Ad Request +-------------------+
| Target Device | --------------------------> | Ad Exchange / DSP |
| (App / Browser) | | (Google, Yandex) |
+------------------+ +-------------------+
| |
| Bid Won & Ad Served | RTB Auction
v v
+------------------+ Evaluates Target Metadata +-------------------+
| 1x1 NOAS iframe | <---------------------------- | Discover-Tech |
| (pqv2.mtrtb[.]com) | (Controlium / RTB)|
+------------------+ +-------------------+
|
| Target Matches (AES-256 Token Verification)
v
+------------------+ Zero-Click Drop +-------------------+
| Exploit Server | ----------------------------> | Candiru Spyware |
| (Targeted Payload| | (COM Hijacking, |
| Only) | | Memory Residue) |
+------------------+ +-------------------+
Operated by Israeli surveillance firm Discover-Tech via its Controlium platform, the system bridges ad exchanges directly with offensive cyber payloads. When an ad slot renders on an end-user device, embedded script tags trigger a call to "NOAS" (getAd) servers (pqv2.mtrtb[.]com). The server evaluates device fingerprints, IP addresses, and Mobile Advertising IDs (MAIDs) in milliseconds. If the incoming profile matches a targeted individual, it injects a 1x1 tracking iframe pointing to exploit infrastructure; if it doesn't, it returns an innocuous 0-byte payload.
The resulting zero-click exploits dropped a Windows DLL attributed to mercenary spyware vendor Candiru (Integrity Labs), using COM hijacking for persistence and 48-byte XOR obfuscation. High-profile compromises spanned dozens of devices at Al Jazeera, while ad impressions tied to the infrastructure surfaced inside government networks across NATO, the UN, the ICC, and the IAEA.
Analysis of 4 billion exposed Apache Druid records revealed that Discover-Tech submitted 3.77 billion bids—winning 441.9 million impressions across 75,373 mobile apps and 120,768 websites via upstream DSPs including Google Ad Manager (24.7%) and Yandex (67.4%). Worse, threat group DARK CARTOGRAPHER combined WebRTC IP leaks with telecommunication carrier "zero-click login" HTTP headers (X-MSISDN) across 30+ countries to directly bridge ephemeral MAIDs to physical phone numbers.
Exposed Ingest & Targeting Scale:
├── 4.0B Exposed Apache Druid analytics records
├── 3.77B Bids submitted across public ad exchanges
├── 441.9M Impressions won across mobile apps and websites
│ ├── Yandex (67.4%)
│ ├── Google Ad Manager (24.7%)
│ └── Beeswax, Vungle, Eskimi (~7.9%)
└── 3.63M Filtered targets routed to profiling endpoints
The reveal sparked fierce backlash across the security community. Discover-Tech had operated for six years as an IAB Europe-approved vendor under the GDPR Transparency & Consent Framework (TCF ID 778), underscoring that ad-tech vendor vetting is built on toothless self-attestation. Furthermore, the exploit delivery employed stringent operational security: payloads were gated behind single-use AES-256 encrypted tokens (?q=...) linked to the winning bid IP, blinding commercial threat crawlers and sandboxes. Third-party ad SDKs and webview iframes must now be treated as untrusted network perimeters—making gateway DNS sinks and MDM-enforced MAID zeroing mandatory corporate defense strategies.
EAP770: When a Corrupted ISO Requires an LPE Kernel Exploit
What starts as routine network troubleshooting often spirals down the low-level rabbit hole. While diagnosing corrupted Linux ISO downloads failing with bad record mac TLS and SSH errors over a Wi-Fi 7 TP-Link EAP770 access point, an engineer uncovered a nasty non-cache-coherent DMA race in Qualcomm's Linux network offload driver (qca-nss-dp).
Packet captures revealed 1440-byte frames containing exactly 64 bytes of zeroed-out, corrupted payload—the exact size of an ARM64 cache line. Crucially, the packets arrived with valid TCP checksums.
Normal Path:
[Host Memory] ---> [Cache Flush] ---> [NIC DMA Read] ---> [Payload Intact]
Corrupted Qualcomm Path:
[Dirty Cache Line]
| (Flushed late to DRAM)
v
[DRAM Corrupted] ---> [PPE Hardware Checksum Calc] ---> [Valid Checksum, Bad Data!]
In edma_tx.c, the fast-transmit path cleaned only skb_headlen bytes before marking buffers with skb->fast_recycled = 1. In edma_rx.c, the refill path spotted this recycling flag and omitted cache invalidation (dmac_inv_range_no_dsb). Dirty cache lines were subsequently flushed to DRAM after network DMA had populated payloads. Because Qualcomm’s hardware Packet Processing Engine (PPE/SFE) calculated checksums in silicon after the corrupt data hit the pipeline, the corrupted payload received a legitimate checksum, blinding transport-layer verification.
Because TP-Link designated the hardware End-of-Life on factory firmware without ever shipping a single patch, fixing the corruption required toggling two vendor sysfs flags (echo 0 > /sys/sfe/l2_feature and ppe_rfs_feature). However, the device's administrative SSH shell is restricted to an unprivileged account (uid 1).
The author chained an exploitation sequence against the access point's Linux 5.4.164 kernel:
- Triggering CVE-2022-0185: Exploited an integer underflow in
legacy_parse_paramvia unprivilegedfsopento corrupt a 4KB kmalloc slab. - KASLR & Ops Leak: Sprayed
keystructs and piped allocations to leakanon_pipe_buf_ops. - Control Hijacking: Forged a fake
pipe_buf_operationsvtable targetinggeneric_pipe_buf_confirmto bypass panic checks during page releases. - Root Execution: Scanned the kernel's
.datasection to overwritemodprobe_path, pivoting to arbitrary root command execution to disable the hardware offload engines.
While turning off offload incurred an 8% throughput penalty (dropping from 103.4 to 95.4 MiB/s), it eliminated payload corruption across all streams. The post highlighted two systemic realities: hardware checksum offloads break the end-to-end guarantees of TCP, and proprietary vendor abandonment cycles continue to transform high-end networking hardware into fragile, vulnerable appliances.
Libreboot 26.10: UEFI Modernization Meets an Anti-AI Mandate
Libreboot released version 26.10 ("Ominous Olive"), introducing a substantial architectural overhaul alongside an aggressive stance on open-source code provenance. The project has formally standardized on MrChromebox's EDK2 UEFI payload across all 64-bit x86 targets, phasing out traditional bare-metal coreboot GRUB and SeaBIOS implementations.
Traditional Libreboot Execution:
[Hardware Reset] ---> [coreboot] ---> [SeaBIOS / GRUB Payload] ---> [Legacy MBR Boot]
Libreboot 26.10 Modern Architecture:
[Hardware Reset] ---> [coreboot] ---> [EDK2 UEFI Payload] ---> [UEFI / UKI / Windows / Linux]
|
+---> [Embedded GRUB EFI]
+---> [CSMWrap / MemTest86+]
By embedding EDK2 directly onto flash chips (≥4MB), Libreboot brings native Windows 10/11 compatibility, Linux Unified Kernel Images (UKIs), and standard NVRAM boot targets to open-firmware laptops and workstations without proprietary binary blobs. The release also integrates support for 233 Intel and AMD Chromebooks via an automated profile importer, restores native AMD AGESA initialization for Fam15h motherboards (fixing multi-socket memory training and broken IOMMU mappings on boards like the ASUS KGPE-D16), and introduces SMMSTORE runtime persistence.
However, the technical upgrade was overshadowed by the release of AGENTS.md—a project-wide zero-tolerance policy against AI-generated code. The policy bans all contributions written or assisted by LLMs, citing unverified intellectual property origins, potential license laundering, and the risks of subtle low-level hardware bugs introduced by statistical code generation.
While parts of the community argued that automated analysis tools (such as the Linux kernel’s Sashiko project) have proven their utility, project maintainers stood firm: in bare-metal boot firmware where debugging requires hardware oscilloscopes and SPI flashers, the opacity of generative slop represents a critical architectural liability.
Liquid AI’s LFM2.5-24B-A2B: Hyper-Sparse Agents on the Edge
Liquid AI debuted LFM2.5-24B-A2B, a 24-billion parameter Mixture-of-Experts (MoE) model engineered for sub-second agentic reasoning on edge hardware. Utilizing extreme routing sparsity, the architecture activates only 2 billion parameters per token.
The model employs 40 layers, 64 experts per MoE block with top-4 routing, combining gated short convolutions with Grouped Query Attention (GQA) over a 131,072-token context window. This architecture allows it to deliver exceptional local throughput:
- 184 tok/s on an Apple M5 Max
- 114 tok/s on an AMD Ryzen AI Max+ 395
- >21,000 tok/s aggregate batched throughput on an NVIDIA B300 (via SGLang BF16)
LFM2.5-24B-A2B Layer Topology
+----------------------------------------------+
| Gated Short Convolutions + GQA |
+----------------------------------------------+
|
v
+----------------------------------+
| MoE Routing Logic (Top-4) |
+----------------------------------+
/ | | \
v v v v
[Exp 03] [Exp 18] [Exp 41] [Exp 62] ... (64 Total)
\ | | /
+------+--------------+-------+
| (2B Active Params)
v
+----------------------------------------------+
| Output Token |
+----------------------------------------------+
Trained across 33 trillion tokens and refined through Multi-Domain On-Policy Distillation (MOPD) inside execution sandboxes, LFM2.5-24B-A2B outperforms larger edge architectures on tool-interaction benchmarks like AutomationBench (27.69 vs. Qwen3.6-35B-A3B’s 8.72) and ToolSandbox (80.71 vs. 75.02).
Evaluation Matrix: Edge Model Tool Execution vs. Code Generation
┌─────────────────────────┬──────────────────┬──────────────────┐
│ Benchmark │ LFM2.5-24B-A2B │ Qwen3.6-35B-A3B │
├─────────────────────────┼──────────────────┼──────────────────┤
│ AutomationBench │ 27.69 │ 8.72 │
│ ToolSandbox │ 80.71 │ 75.02 │
│ LiveClaw │ 48.09 │ 38.19 │
│ SWE-Bench Verified │ 51.00 │ 66.20 │
└─────────────────────────┴──────────────────┴──────────────────┘
Real-world community testing, however, revealed a pronounced architectural trade-off. While the model handles tool dispatch loops rapidly, its code reasoning capabilities lag behind (scoring 51.00 on SWE-Bench Verified against Qwen's 66.20). Early adopters noted that without supervision, the model can enter high-speed hallucination loops—humorously summarized on Hacker News as: "It is very fast, but also very stupid; it will delete prod at 250 t/s." The emerging consensus points toward a tiered agent architecture: using ultra-sparse models as fast, local tool routers that defer complex code synthesis and high-entropy reasoning tasks to heavier cloud-based models.
Low-Level Engineering: Voxel Rack Emulation, C++26 Contracts, and Double Stacks
Beyond headline security and machine learning developments, several foundational systems engineering breakthroughs were unveiled today:
- Emulating Data Centers on a Mini-PC (Oxide Voxel): Oxide Computer introduced Voxel, an open-source platform capable of virtualizing an entire Oxide rack on a single AMD Ryzen mini-PC. Rather than relying on high-level software mocks, Voxel boots the stock, unmodified Omicron control plane and Helios OS. Compute sleds run inside Propolis (bhyve) hypervisors, switching fabrics are handled by a P4 SoftNPU, and physical service processors run within cycle-accurate Cortex-M7/M33 emulators (
sp-emu). To maintain stability on hardware running slower than bare metal, the platform detaches RTOS timers from real-world time: the Hubris 1ms SysTick fires strictly according to CPU instruction counts rather than system wall-clock cycles. - C++26 Formalizes Hoare Logic (Contracts P2900): The C++ Standards Committee advanced contract assertions into the C++26 working draft. Developers can now express pre- and post-conditions directly within declarations:
The compiler supports four distinct evaluation modes:int clamp(int val, const int low, const int high) pre(low <= high) post(r: r >= low && r <= high);ignore,observe,enforce, andquick-enforce. The standard explicitly mandates that contracts are diagnostic metadata rather than control-flow mechanisms; arguments within contract predicates are implicitly treated asconstto prevent compiler side effects from altering production logic. - Double-Headed Stack Allocators: Former game engine architects resurfaced an elegant memory optimization pattern: the double-headed bump allocator. By positioning a single contiguous buffer where Stack 0 grows upward from index
0and Stack 1 grows downward from indexN, two independent memory lifecycles can coexist without fragmentation. An OOM condition only triggers when the two pointers cross (tos0 > tos1). Commonly used in streaming systems—such as pairing volatile level geometry with static character meshes—the model provides O(1) alloc/free cycles without the latency variance or memory overhead of traditional general-purpose dynamic heap managers.
Key Takeaways for Builders
- Audit Third-Party Ad and Analytics Integrations: The Citizen Lab disclosure confirms that RTB exchanges and client-side ad SDKs serve as active vectors for zero-click remote code execution. Network perimeters must enforce strict DNS and IP blocklists against real-time bidding endpoints on enterprise-managed devices.
- Beware Checksum Offloading on Asymmetric SoCs: If developing for or deploying on non-cache-coherent architectures (common in embedded ARM platforms), never assume transport-layer TCP checksums guarantee data integrity. Hardware offload units compute checksums downstream from memory, making application-layer cryptographic validation (TLS/SSH) your only true defense against silent bit-rot and dirty cache flushes.
- Decouple Agent Tool-Dispatch from Reasoning: Hyper-sparse MoE models like LFM2.5 prove that sub-second tool routing can run efficiently on edge hardware. However, their lower performance on complex semantic benchmarks suggests production agent architectures should employ small, fast models for environment orchestration, while delegating complex code modification to specialized models.