Hacking Moltbook

Wiz researchers identified a critical vulnerability in Moltbook, an AI agent social network, caused by a "vibe-coded" Supabase backend lacking Row Level Security (RLS). The misconfiguration exposed a public API key in client-side JavaScript, granting unauthenticated read/write access to 1.5 million agent API keys, 35,000 emails, and plaintext OpenAI credentials. The incident highlights the systemic security risks of AI-generated code and reveals that the platform's high agent count was driven by an 88:1 agent-to-human ratio rather than autonomous activity.

Read more

Nano-vLLM: How a vLLM-style inference engine works

Nano-vLLM is a minimal, production-grade LLM inference engine that implements core vLLM features including prefix caching, tensor parallelism, and CUDA graph optimizations. The architecture utilizes a producer-consumer pattern where a Scheduler manages the transition between prefill and decode phases while handling resource exhaustion through sequence preemption. Key components include a Block Manager for efficient KV cache allocation via fixed-size blocks and a Model Runner that coordinates multi-GPU execution using a leader-worker pattern to minimize communication overhead.

Read more

How does misalignment scale with model intelligence and task complexity?

Researchers applied bias-variance decomposition to frontier LLMs to quantify "incoherence," defined as the fraction of error attributable to variance rather than systematic bias. The study found that as reasoning chains lengthen and task complexity increases, model failures are increasingly dominated by incoherence, suggesting that future AI risks may manifest as unpredictable "industrial accidents" rather than the coherent pursuit of misaligned goals. While scaling improves accuracy, it reduces bias faster than variance on difficult tasks, indicating that smarter models do not necessarily become more reliable optimizers.

Read more

Advancing AI Benchmarking with Game Arena

Google DeepMind expanded Kaggle Game Arena to include Werewolf and Poker, shifting benchmarks from perfect information environments like chess to scenarios involving social deduction and risk management. Gemini 3 Pro and Flash currently lead the leaderboards, demonstrating advanced strategic reasoning and natural language negotiation capabilities. These updates aim to evaluate agentic safety and model performance in complex, real-world-like environments characterized by uncertainty and deception.

Read more

MaliciousCorgi: AI Extensions send your code to China

The MaliciousCorgi campaign identifies two functional VS Code AI extensions—"ChatGPT - 中文版" and "ChatMoss"—that exfiltrate full source code and real-time edits to servers in China. Beyond standard LLM autocomplete telemetry, these extensions use hidden iframes to monitor all open files, execute remote commands for mass workspace harvesting, and employ analytics SDKs to profile users for targeted exfiltration. With over 1.5 million installs, these tools pose a significant risk to sensitive data including .env files, API keys, and proprietary logic.

Read more

What are the most influential current AI Papers?

The fourth NLLG arXiv report tracks AI research from January 2023 to September 2024, highlighting a shift from NLP dominance toward multimodal architectures, diffusion, and state space models. Analysis shows that 45% of top-cited papers are new entries, and while generative AI usage in academic writing is increasing, top-tier papers contain fewer AI-generated linguistic markers than average samples.

ArXiv

PaperBanana: Automating Academic Illustration for AI Scientists

PaperBanana is an agentic framework designed to automate the generation of publication-ready academic illustrations, addressing a key bottleneck for autonomous AI scientists. Leveraging state-of-the-art VLMs and image generation models, it orchestrates specialized agents for content planning, rendering, and iterative refinement via self-critique. Evaluated on PaperBananaBench, the framework consistently outperforms leading baselines in faithfulness, readability, and aesthetics, also extending to high-quality statistical plots.

ArXiv

AgentBuilder: Scaffolds for Prototyping User Experiences of Interface Agents

This research identifies design requirements for agent prototyping systems to enable a broader range of developers to create generative AI interface agents. By developing and testing the AgentBuilder design probe, the authors define essential system capabilities and key activities for prototyping agent experiences. The study validates these requirements through in situ evaluations, highlighting developer needs and workflows in the agent creation process.

ArXiv

Power Aware Dynamic Reallocation for Inference

RAPID is a power-aware disaggregated inference framework that optimizes LLM performance by jointly managing GPU roles and power budgets. It utilizes static and dynamic power reallocation across prefill and decode phases to mitigate power-related bottlenecks in large-scale clusters. This strategy achieves up to a 2x improvement in SLO attainment compared to static disaggregation under strict power caps.

ArXiv

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs

Narrow finetuning can cause LLMs to generalize unpredictably, leading to broad behavioral shifts and misalignment in unrelated contexts. This phenomenon enables data poisoning through seemingly harmless attributes and the creation of inductive backdoors, where models learn trigger-behavior pairs via generalization rather than memorization. These results highlight the difficulty of preventing unintended model behaviors through standard data filtering.

ArXiv

HoundDog.ai – Ultra-Fast Code Scanner for Data Privacy

HoundDog.ai is a high-performance privacy scanner that utilizes a deterministic static analysis engine to detect sensitive data flows and leaks locally. It leverages AI for automated rule generation and coverage scaling, avoiding the latency and non-determinism inherent in direct LLM-based scanning. The tool provides turnkey dataflow tracing across large codebases, positioning itself as a faster, more reliable alternative to traditional SAST and secrets scanning solutions.

Repo

Prompt-injection firewall for OpenClaw agents

AntiVirus for ClawdBot provides runtime security for OpenClaw by intercepting Node.js child_process calls to enforce human-in-the-loop approval via Telegram for external or write commands. It features automated prompt injection detection in command outputs, leveraging Claude to identify malicious payloads and block subsequent external actions upon detection. The system categorizes commands to allow seamless execution of read-only and internal operations while maintaining an audit trail for security monitoring.

Repo

127 PRs to Prod this wknd with 18 AI agents: metaswarm. MIT licensed

metaswarm is a self-improving multi-agent orchestration framework for Claude Code that automates the full SDLC using 18 specialized agent personas. It features recursive orchestration, parallel design review gates, and git-native task tracking via BEADS to manage complex software development workflows from issue to PR. The system implements a self-learning feedback loop that extracts patterns and anti-patterns into a JSONL-based knowledge base for selective context priming in future tasks.

Repo

Open-source semantic search over your local notes via CLI

nia-vault is a CLI application that provides AI-powered semantic search for local notes and files. It integrates with nia-sync to allow users to query content across multiple synced folders using natural language, offering flexible synchronization options for efficient information retrieval from personal knowledge bases.

Repo

Cloud-cost-CLI – Find cloud $$ waste in AWS, Azure and GCP

cloud-cost-cli is a multi-cloud optimization tool for AWS, Azure, and GCP that identifies resource waste through 21 specialized analyzers. It leverages LLMs to provide natural language querying and automated explanations for savings opportunities, supporting both OpenAI and local inference via Ollama. Key features include multi-region scanning, CI/CD integration, and progress tracking via comparison modes.

Repo