Today’s edition covers radical departures in CPU silicon design, hardware fault injection that liberates enterprise GPUs from crypto-mining silicon, and a decisive architectural pivot toward deterministic, hierarchical AI agent infrastructure.
Silicon Re-Architectures: From Post-Out-of-Order CPUs to Glitched GA100 Dies
Tendrils Raises $30M to Ditch the Sequential CPU
Benchmark has led a $30M Series A into Tendrils, a hardware startup aiming to solve the sequential compute bottleneck by stripping out decades of general-purpose architectural baggage: out-of-order execution logic (OoO), reorder buffers (ROB), and hardware-enforced cache coherency. In their place, Tendrils is building a processor with thousands of lightweight parallel cores designed specifically to execute Vine, an open-source language based on interaction nets and parallel graph reduction.
The systems engineering community met the announcement with healthy historical skepticism. Veterans quickly drew parallels to past architectures that attempted to replace sequential caches with massive spatial parallelism—such as the INMOS Transputer, Tilera, GreenArrays, and the Sony/Toshiba/IBM Cell Broadband Engine, whose notorious 256KB local store bottleneck crippled developer productivity. Without OoO hardware or aggressive simultaneous multithreading (SMT) to mask memory latency, instruction execution units stall during non-predictable memory fetches. Unless Tendrils demonstrates how graph reduction circumvents cache locality fragmentation and pointer-chasing latency at the hardware level, this ambitious architecture risks repeating the classic compilation-friction traps of unconventional silicon.
Unlocking an A100 from an Ethereum Mining Card via Voltage Fault Injection
Hardware security researcher Kate published a technical breakdown demonstrating how to bypass TSMC eFuse protections on NVIDIA CMP 170HX mining cards using targeted voltage fault injection (VFI). The CMP 170HX houses a physical GA100 die and between 40GB to 80GB of HBM2e memory across up to six stacks, but NVIDIA artificially binned and throttled the cards to 70 Streaming Multiprocessors (SMs) and crippled PCIe bandwidth via hardware fuses.
+---------------------------------------------------------------+
| CMP 170HX (Artificially Restricted GA100 Silicon) |
| - 70 SMs Active | PCIe Gen1/2 Throttled | HBM2e Capped |
+---------------------------------------------------------------+
|
[STM32G4 Microcontroller Intercept]
Monitors NVVDD rail; Triggers at 225 mV
Glitching MP2888B VRM PWMVID lines
|
v
+---------------------------------------------------------------+
| TSMC eFuse Sensing Disrupted: SBFE cleared to 0 |
| -> Arbitrary Write Access to Fuseless Fuse (FF) Registers |
| -> Override Factory Restriction Records |
+---------------------------------------------------------------+
|
v
+---------------------------------------------------------------+
| Unlocked Enterprise Accelerator |
| - 124 SMs Active (7,936 Cores @ 1410 MHz) |
| - 85.86 TFLOP/s FP16 (~1.18x over retail A100 108-SM base) |
| - PCIe Gen4 x16 & NVLink NV12 Fully Operational |
+---------------------------------------------------------------+
By intercepting the PWMVID communication between the GPU and the MP2888B VRM controller using an STM32G4 microcontroller, Kate injected precise voltage drops as NVVDD crossed 225 mV during boot. This glitch disrupted the sensing of control rows 0 and 1, clearing OPT_SECURED_BURN_FEATURE_ENABLE (SBFE=0) and opening arbitrary write access to the Fuseless Fuse (FF) region (rows 222–487).
The result? Overriding the factory configuration unlocked 124 SMs (7,936 cores) at 1,410 MHz, yielding 85.86 TFLOP/s of FP16 compute (beating a stock retail 108-SM A100's 72.49 TFLOP/s), enabling PCIe Gen4 x16, and activating NVLink. The hack illustrates a profound security paradox in modern silicon binning: because replacement fuse records override previous ones in sequential chained instruction designs, achieving write access allows an attacker to relax hardware constraints rather than permanently brick the chip.
Black Ops II Statically Recompiled for ARM64 macOS
In an extraordinary systems preservation project, developer adrbn released bo2-mac, a native macOS port of Call of Duty: Black Ops II that completely bypasses Wine, CrossOver, and Apple’s Rosetta 2 translation layer.
The toolchain lifts roughly 150,000 32-bit x86 functions into LLVM intermediate representation using Ghidra and Remill v6.0.1. Custom LLVM passes fold guest register stores, map guest pointers into a direct 4 GiB virtual address space, and map x86 lock instructions to native ARM64 atomics before compiling with clang -O2. Graphics are handled by a custom fake-COM Direct3D 11-to-Metal translation layer, converting DXBC bytecode to MSL via HLSLcc.
To handle obfuscated copy protection and dynamically unpacked code without decompiling them ahead of time, the runtime embeds an in-process Unicorn x86 interpreter fallback. On an entry-level base 2020 M1 MacBook Air, the game runs at a stable 50–60 FPS, demonstrating how ahead-of-time binary lifting combined with targeted runtime emulation can resurrect legacy x86 software on modern ARM architectures.
The New AI Control Plane: Deterministic Gates & Hierarchical Agents
Gemini Decisions API: Bounded Multimodal Routing at 50ms Latency
Google DeepMind unveiled the Gemini Decisions API (gemini-decision-1), a multimodal endpoint engineered specifically for closed-set classification that bypasses autoregressive token-by-token text generation entirely.
Operating over text, audio, video, and documents up to 256K tokens, the endpoint evaluates inputs in a single forward pass, delivering compute latencies of 50–100 ms (150–250 ms end-to-end) with flat response curves across 1 to 100 questions. It introduces three core primitives: noul (calibrated Yes/No probabilities), choice (categorical selection across up to 255 discrete options), and score (weighted rubric scoring).
Priced at $0.10 per 1M input tokens with $0 output billing, this architecture replaces the fragile practice of prompting large models for structured JSON routing tokens. Developers gain properly calibrated probability distributions without the miscalibration often introduced by RLHF/RLVR alignment layers, offering a low-latency, deterministic gating mechanism for multi-agent systems.
The Rise of Hierarchical Agent Execution: shadcn/improve and Meta Muse Ember
The developer ecosystem is increasingly abandoning the concept of a single model that plans and executes code simultaneously. Two releases this week signal a shift toward hierarchical, multi-tier agent orchestration:
shadcn/improve: An open-source auditing harness that uses expensive frontier reasoning models to perform deep codebase reconnaissance across security, DX, and correctness. Crucially, the planning agent never writes code directly. Instead, it outputs self-contained Markdown specifications with inlined context, concrete test assertions, and strictSTOPboundaries. These plans are then handed off to inexpensive worker models executing inside isolatedgit worktrees.- Meta Muse Ember: Meta launched an ultra-low-cost, high-throughput model variant explicitly tailored for high-volume worker orchestration, competitive with Gemini Flash-Lite. Ember is designed for the execution end of the pipeline—running repetitive tool calls, data parsing, and test-driven edits where frontier reasoning models represent massive overkill in cost and latency.
+-----------------------------------------------------------+
| Frontier Reasoning Model |
| (Claude 3.7 / o1 / GPT-4.5 Tier) |
| - Deep Architectural Reconnaissance |
| - Generates Self-Contained Markdown Specs & Verification |
+-----------------------------------------------------------+
|
v
+-----------------------------------------------------------+
| Low-Cost Execution Fleet |
| (Meta Muse Ember / Flash-Lite / Qwen) |
| - Isolated Git Worktrees |
| - Deterministic Tool Loops via MCP (e.g., 1ToolRing) |
| - Run Unit Tests Until Assertion Exits 0 |
+-----------------------------------------------------------+
This pattern matches experiments like 1ToolRing, an infrastructure project that consolidates 3,051 paid third-party APIs behind a unified Model Context Protocol (MCP) proxy. By decoupling expensive planning from execution and routing it through dynamic balance checks and structured schemas, developers are finally turning unpredictable autonomous loops into cost-bounded engineering workflows.
LiDAR Vision Agents Unearth Classic Maya Architecture
Demonstrating the real-world utility of multimodal agent pipelines, independent researcher Pietro Miotti deployed an array of 48 parallel LLM vision agents to analyze 1,200 km² of public NASA G-LiHT airborne LiDAR data over Campeche, Mexico.
After peeling back the thick 10–17-meter rainforest canopy to build high-resolution Digital Elevation Models (DEMs), the agent fleet was prompted with archaeological heuristics to distinguish anthropogenic rectangular mounds, aguadas (reservoirs), and plazas from natural karst sinkholes. Within 72 hours, the system flagged an unmapped 8-to-12-meter-tall oval pyramid and acropolis complex. The discovery was subsequently verified against official registers by Mexico's National Institute of Anthropology and History (INAH), demonstrating that scaled, multimodal vision models can execute non-trivial spatial pattern recognition without bespoke CNN segmentation architectures.
Apple "Harbor" & Local Inference via Thunderbolt 5
Apple is reportedly finalizing a native macOS application codenamed Harbor under the Claris banner, aimed at turning desktop Apple Silicon hardware into zero-configuration local AI servers.
The software’s headline capability is automated inference clustering over Thunderbolt 5. By linking Macs via high-speed physical interconnects, Harbor aggregates unified memory pools into a single logical address space for open-weight models. While hardware enthusiasts are celebrating the prospect of pooling multiple Mac Studios to run 70B+ parameter models, infrastructure engineers note that TB5’s 80–120 Gbps bandwidth is still vastly slower than intra-chip unified memory buses (400–800 GB/s on Max and Ultra chips). Tensor parallelism across nodes will face substantial interconnect bottlenecks unless Harbor partitions models using pipeline parallelism.
Cloud Emulation and Network Geolocation Realities
AWS Open-Sources Its Official Cloud Emulator—Written by Agents
In an effort to provide a local integration testing ground and challenge LocalStack, AWS open-sourced aws-emulator, a local Rust-based mock for core AWS APIs. Built with crates generated directly from AWS Smithy interface models, the project supports an Embedded Mode where in-process Rust integration tests can route SDK calls without network socket overhead.
Service Emulation Pass Rate
---------------------------------------
Data Firehose [██████████░░░░] 77%
KMS [█████████░░░░░] 70%
S3 [████████░░░░░░] 62%
Lambda [███████░░░░░░░] 53%
EC2 [█████░░░░░░░░░] 40%
CloudFormation [████░░░░░░░░░░] 36%
Athena [███░░░░░░░░░░░] 22%
The release ignited heated discussion when AWS revealed that autonomous AI coding agents generated every line of the emulator across more than 29,000 commits. A close inspection of the codebase explains why: service fidelity is wildly inconsistent. Test pass rates range from 77% for Kinesis Firehose down to 53% for Lambda, 40% for EC2, and 22% for Athena. Crucially, the security documentation warns that the emulator does not validate IAM policies, KMS key configurations, or cross-account trust boundaries. While useful as a syntactic mock for happy-path developer loops, it cannot replace genuine staging environments for runtime security invariants.
Speed-of-Light Latency Exposes AFRINIC IPv4 Leakage
A collaborative network measurement study (WHEREIS, presented at ACM IMC 2026) by IPinfo and academic researchers proved that 2.43% of AFRINIC’s allocated IPv4 address space actually answers from outside the African continent.
Using simple physics—the speed of light in optical fiber (~200 km per millisecond)—researchers proved physical distance via round-trip ping times (RTT). A 0.28 ms ping from a probe in Lisbon to an IP delegated to the Democratic Republic of the Congo confirmed the server was physically located within 30 kilometers of the Portuguese capital.
The study found that roughly 81% of these displaced addresses (over 460,000 active IPs) belong to four large blocks assigned to Cloud Innovation Ltd, a entity registered in the Seychelles that has been embroiled in legal battles with AFRINIC in Mauritius since 2021. The operational takeaway is stark: Regional Internet Registry (RIR) registration data and static GeoIP databases increasingly reflect corporate ownership and legal shell companies rather than physical network topology, creating major blind spots for geolocation-based compliance, fraud engines, and edge routing.
Endpoint Security & Local Privacy
Hardening SSH Keys with the macOS Secure Enclave
A guide by Gatezh brought renewed attention to binding OpenSSH keys directly to the macOS Secure Enclave via CryptoTokenKit. By generating hardware-bound NIST P-256 identities (sc_auth create-ctk-identity) and creating FIDO2 handles (sk-ecdsa-sha2-nistp256@openssh.com), private key material becomes physically non-exportable and requires Touch ID authorization for every cryptographic signature.
# Generate hardware-bound Secure Enclave identity requiring biometric auth
sc_auth create-ctk-identity -l "ssh-MacBook" -k p-256-ne -t bio
# Link to OpenSSH via system keychain provider
ssh-keygen -w /usr/lib/ssh-keychain.dylib -K -N ""
export SSH_SK_PROVIDER=/usr/lib/ssh-keychain.dylib
While this provides airtight protection against info-stealing malware extracting ~/.ssh/id_* files from disk, developers pointed out real-world friction: invoking Touch ID on automated Git loops (such as multi-commit interactive rebases or agentic tools polling remote repositories) triggers severe prompt fatigue. Furthermore, hardware-enclave keys break headless automation, CI workflows, and remote SSH agent-forwarding pipelines, requiring teams to maintain a strict separation between human-interactive developer keys and decoupled service tokens.
The Smart TV Privacy Void: Audio Hashes and Active Network Reconnaissance
A sweeping investigation across eleven major television and streaming platforms (including LG, Samsung, Roku, Amazon, and Google TV) laid bare the aggressive telemetry techniques powering Automatic Content Recognition (ACR).
The findings dismantle common assumptions about content tracking: manufacturers do not need to transmit raw video frames to monitor your viewing habits. Instead, LG’s webOS extracts real-time acoustic perceptual fingerprints directly from the internal digital sound processor (DSP)—bypassing speakers entirely—to cross-reference against commercial ad and media databases. This fingerprinting operates indiscriminately across broadcast channels, native apps, local Plex/Jellyfin streams, and connected HDMI consoles.
Coupled with admissions that Google TV and LG actively run local network scans (via mDNS, SSDP, and Matter) to profile other devices on your home LAN, consumer televisions have effectively evolved into advertising and telemetry terminals subsidized by low hardware prices. The engineering consensus remains unequivocal: isolate smart TVs on an airgapped IoT VLAN with WAN-level DNS blocking, or treat them strictly as "dumb" monitors paired with privacy-hardened external playback devices.
Yogurt: Private, Botless Meeting Transcription
Addressing enterprise surveillance fatigue, Frigade open-sourced Yogurt, a local-first alternative to AI note-takers like Granola.
Built using Electron, node-llama-cpp, and local Whisper bindings across macOS, Windows, and Linux, Yogurt avoids the intrusive convention of inviting third-party recording bots to video calls. Instead, it captures internal system audio and local microphones directly via system loopback.
To solve the speaker diarization challenge without the compute overhead of heavy speaker-embedding models like Nemotron, Yogurt inspects the operating system's native accessibility trees to dynamically read participant names as their speech indicators highlight in the Google Meet or Zoom interfaces. It is an effective engineering workaround that keeps both raw audio and conversation context strictly on local device memory.