Google broke its promise to me – now ICE has my data

Google violated its commitment to notify users before disclosing data to law enforcement by providing account metadata to ICE without warning. The EFF has filed complaints with state Attorneys General, arguing that this failure to allow legal challenges constitutes a deceptive trade practice. This incident highlights the privacy risks associated with centralized data collection and the use of session logs and IP addresses to construct detailed surveillance profiles.

Read more

Open Source Isn't Dead

Cal.com is transitioning to a closed-source model, citing the rise of low-cost, AI-automated vulnerability discovery. Strix argues that closing source code is an ineffective defense against modern AI agents capable of black-box and grey-box testing on live endpoints. They contend that security through obscurity cannot scale against automated exploitation and advocate for integrating autonomous AI security agents directly into the CI/CD pipeline to provide continuous, proactive defense.

Read more

Google Gemma 4 Runs Natively on iPhone with Full Offline AI Inference

Google Gemma 4 now supports native, offline inference on iPhone via the Google AI Edge Gallery, utilizing the device's GPU for low-latency execution. While the 31B variant benchmarks similarly to Qwen 3.5, the E2B and E4B models are specifically optimized for mobile memory and thermal constraints. This release facilitates edge AI deployment for privacy-critical enterprise and healthcare applications without cloud dependency.

Read more

Elevated errors on Claude.ai, API, Claude Code

Between April 1 and April 15, 2026, Claude services experienced frequent disruptions, primarily impacting authentication across Claude.ai, Claude Code, and the API. Notable incidents included recurring elevated error rates for Sonnet 4.6 and Opus 4.6 models, degraded performance for Vaults and Connectors, and outages affecting workspace creation and admin API endpoints. Most issues were resolved within hours, though login stability remained a persistent challenge throughout the period.

Read more

AI-assisted cognition endangers human development?

AI-assisted cognition risks inducing "AI-skew" and cognitive inbreeding by anchoring human thought to the static inductive biases of LLM base models. Because post-training often fails to fully integrate new cultural or factual shifts into a model's internal representations, population-level reliance on a few dominant models narrows the diversity of the "Dynamic Dialectic Substrate." Mitigating this requires "cognition hygiene" strategies, such as utilizing diverse base models, leveraging web search to bypass direct model patterns, and prioritizing human-to-human dialectic processes.

Read more

A Survey of Workflow Optimization for LLM Agents

This survey formalizes LLM-based workflows as Agentic Computation Graphs (ACGs), categorizing optimization methods into static and dynamic structures based on when the workflow is determined. It analyzes ACGs across three dimensions—timing, optimization targets, and feedback signals—while distinguishing between reusable templates and run-specific execution traces. The paper advocates for structure-aware evaluation metrics, including graph properties, cost, and robustness, to standardize the development of LLM agent workflows.

ArXiv

Aethon: A reference-based instantiation primitive for stateful AI agents

Aethon introduces a reference-based replication primitive to address the significant latency and memory overhead associated with instantiating stateful AI agents, a shift from traditional stateless LLM inference. Instead of fully materializing agents, Aethon represents instances as compositional views over stable definitions, layered memory, and contextual overlays, enabling near-constant-time instantiation. This approach, leveraging layered inheritance and copy-on-write semantics, decouples creation cost from inherited structure, enhancing scalability and orchestration for production-scale agentic software.

ArXiv

Interpreting Negation in GPT-2: Layer- and Head-Level Causal Analysis

Researchers performed a causal analysis of negation processing in GPT-2 Small using activation patching and ablation on a 12,000-pair dataset. By measuring the Negation Effect Score (NES), they localized negation sensitivity to a specific subset of mid-layer attention heads, primarily within layers 4–6. These findings, validated on the xNot360 benchmark, demonstrate that logical polarity is handled by concentrated internal components rather than being distributed throughout the model.

ArXiv

Lyra 2.0: Explorable Generative 3D Worlds

Lyra 2.0 addresses spatial forgetting and temporal drifting in generative 3D reconstruction by maintaining per-frame geometry for information routing and training with self-augmented histories. This framework enables long-horizon, 3D-consistent video generation for large-scale environment creation, facilitating high-quality feed-forward reconstruction from extended camera trajectories.

ArXiv

Watching TV with the Second-Party: A First Look at ACR Tracking in Smart TVs (2024)

A black-box audit of ACR on Samsung and LG smart TVs reveals that content-based tracking persists across all input methods, including external HDMI sources. While privacy opt-out controls effectively terminate ACR network traffic, the study identifies significant regional discrepancies in tracking implementation between the US and UK.

ArXiv

Does Gas Town 'steal' usage from users' LLM credits to improve itself?

Gas Town is a multi-agent orchestration system for AI coding agents like Claude Code and GitHub Copilot, designed to scale workflows to dozens of agents. It ensures state persistence across agent restarts using git-backed hooks and a structured data ledger called Beads. The architecture features a central AI coordinator (Mayor), ephemeral worker agents (Polecats), a Bors-style merge queue (Refinery), and a three-tier watchdog system for health monitoring and automated recovery. It also supports session history discovery via Seance and federated coordination through Wasteland.

Repo

Libretto – Making AI browser automations deterministic

Libretto is an open-source toolkit that provides coding agents with a live browser and a token-efficient CLI for building and maintaining web integrations. It utilizes LLMs for off-thread snapshot analysis to extract selectors and diagnose failures, minimizing context window overhead for the primary agent. Key capabilities include reverse-engineering site APIs from network traffic, recording user actions for Playwright script generation, and autonomously debugging broken automation workflows.

Repo

Jeeves – TUI for browsing and resuming AI agent sessions

Jeeves is a Go-based TUI for browsing, searching, and resuming AI agent sessions from Claude Code, Codex, and OpenCode. It features regex-enabled search, conversation previews, and the ability to relaunch sessions directly in the agent. The tool leverages the Charm ecosystem to provide a centralized interface for managing local agent history.

Repo

AI support chatbot with RAG and citations – one back end file, no infra

OnCell provides a unified infrastructure for building AI support agents, eliminating the need to manage separate vector databases, file storage, or conversation stores. It offers an isolated runtime environment with built-in vector search and NVMe-backed storage to facilitate RAG with citations and per-customer memory. Developers can deploy a complete backend using a single logic file that integrates with OpenRouter for LLM orchestration.

Repo

Slop-scan – Detect AI code slop patterns in your repo

slop-scan is a deterministic CLI tool that identifies AI-associated "slop" patterns in JS/TS repositories using explainable heuristics rather than authorship detection. It surfaces hotspots like redundant async wrappers and duplicate function signatures, providing normalized metrics to benchmark codebases against mature OSS baselines. The pluggable engine supports custom rules and outputs findings in lint or JSON formats for CI/CD integration.

Repo