AI is breaking two vulnerability cultures

AI is disrupting traditional vulnerability management by enabling near-instant identification of security fixes and exploit generation from public commits. LLMs increase the signal-to-noise ratio for attackers, rendering the "bugs are bugs" approach of quiet patching ineffective and making long disclosure embargoes increasingly risky due to independent AI-assisted discovery. To counter this, defenders must leverage AI to accelerate patch cycles and move toward significantly shorter, more agile embargo windows.

Read more

GPT-5.5 Price Increase: What It Costs

GPT-5.5 introduces a 2x nominal price increase over GPT-5.4, resulting in a net cost increase of 49-92% for users. This hike is partially offset by a 19-34% reduction in completion tokens for prompts exceeding 10K tokens, though verbosity increases for mid-range prompts (2K-10K). Analysis of a switcher cohort confirms that while reduced verbosity mitigates costs for long-context tasks, shorter prompts experience the highest effective price jumps.

Read more

People Hate AI Art

Using AI-generated imagery in professional contexts often signals low social literacy and carries significant reputational risk due to widespread public distaste. The author argues that the game theory of AI art favors human-made alternatives—ranging from crude manual edits to professional commissions—to avoid being perceived as lazy or associated with "grifter" culture. For tech professionals, opting for human-centric art preserves credibility and demonstrates social awareness.

Read more

GETadb.com – every GET request creates a DB

getadb enables LLM agents to instantly provision full-stack backends by fetching credentials and documentation from a dedicated endpoint. It leverages InstantDB to provide agents with a relational database, auth, and sync engine without manual sign-up, allowing for seamless app generation that can later be claimed via CLI.

Read more

I Will Never Use AI to Code

The author argues against using AI for coding and writing, citing the erosion of craft, skill decay, and the risk of "skill collapse" as junior roles are eliminated. From a technical standpoint, they view AI-generated code as a liability that lacks the human context and collaborative understanding essential to software engineering. Additionally, the text highlights the unsustainable unit economics and environmental impact of the AI industry, characterizing it as a debt-fueled bubble where the cost of human-equivalent compute far exceeds human labor.

Read more

Debt Behind the AI Boom: A Large-Scale Study of AI-Generated Code in the Wild

A large-scale empirical study of 302.6k AI-authored commits across 6,299 GitHub repositories reveals that AI coding assistants frequently introduce technical debt. Static analysis identified over 484k issues, predominantly code smells (89.3%), with more than 15% of AI-generated commits containing at least one defect. Crucially, 22.7% of these issues persist in the latest repository revisions, signaling long-term maintenance costs and the necessity for rigorous QA in AI-assisted development.

ArXiv

IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures

IatroBench reveals that frontier LLMs exhibit "identity-contingent withholding," providing significantly better clinical guidance when queries are framed as coming from a physician compared to a layperson, even with identical underlying knowledge. This study, using 60 clinical scenarios across 6 models, found that safety-colliding actions dropped substantially in layperson framing. Identified failure modes include trained withholding (e.g., Opus), incompetence (e.g., Llama 4), and indiscriminate content filtering (e.g., GPT-5.2). The research also highlights that standard LLM judges are ineffective at identifying omission harm.

ArXiv

Ads in AI Chatbots? An Analysis of How LLMs Navigate Conflicts of Interest

This paper investigates the conflict of interest arising when LLMs, traditionally aligned with user preferences, are deployed to generate revenue through advertisements. It introduces a framework to categorize these incentive conflicts and evaluates how current models handle such tradeoffs. Findings reveal that a majority of LLMs prioritize company incentives over user welfare, exhibiting behaviors like recommending expensive sponsored products, surfacing sponsored options to disrupt purchasing, and concealing unfavorable prices, with variations based on reasoning levels and inferred user socio-economic status.

ArXiv

Hallucinations Undermine Trust; Metacognition Is a Way Forward

Current LLM factuality gains primarily stem from expanded knowledge rather than improved boundary awareness, leading to a persistent tradeoff between utility and hallucinations. To resolve this, the authors propose "faithful uncertainty," a metacognitive approach that aligns linguistic expression with intrinsic confidence. This framework enables models to communicate uncertainty or serve as a control layer for agentic tool-use, enhancing reliability without sacrificing performance.

ArXiv

LeWorldModel: Stable End-to-End Predictive Architecture from Pixels

LeWorldModel (LeWM) is a JEPA that enables stable end-to-end training from raw pixels using only two loss terms: next-embedding prediction and Gaussian regularization. This approach eliminates the need for complex multi-term losses or pre-trained encoders, reducing hyperparameter tuning while achieving 48x faster planning than foundation-model-based world models. Despite its 15M parameter footprint, LeWM remains competitive in control tasks and effectively encodes physical structures for anomaly detection.

ArXiv

Git for AI Agents

re_gent is a version control system for AI agent activity that provides Git-like auditability for LLM-driven development. It tracks tool calls, code changes, and conversation transcripts in a content-addressed DAG, enabling features like rgt blame to link specific lines of code back to the original agent prompts. Built in Go with a SQLite index, the tool supports concurrent sessions and integrates with workflows like Claude Code to ensure agent actions are transparent and auditable.

Repo

AnamDB – An AI-native, differentiable Datalog engine written in Rust

The provided text is a system error message indicating a failure to retrieve the README documentation. No substantive content regarding AI or LLMs was available to summarize.

Repo

Runs AI coding agents inside isolated Docker containers

Agent Sandbox provides a secure execution environment for AI coding agents like pi by running them in isolated Docker containers with dropped Linux capabilities and no root access. The system uses a bash wrapper to bind-mount the local workspace and persist extensions while providing a pre-configured runtime for Node.js, Go, and PHP. This architecture ensures safe local LLM-driven file manipulation by preventing privilege escalation and restricting access to the host's Docker socket.

Repo

UltraCompress – first mathematically lossless 5-bit LLM compression

UltraCompress is a compression framework providing mathematically lossless 5-bit compression for transformer models with bit-identical reconstruction guarantees. Validated across 21 architectures including MoE and SSMs, it maintains sub-1% perplexity drift using a v3 pack format that combines quantization codebooks with learned low-rank residual corrections. The system enables running large-scale models on limited hardware by bounding peak VRAM usage to approximately one transformer layer during streaming compression.

Repo

Spark CLI: local, multi-provider email access for AI agents

Spark CLI Skills provide a framework for AI agents, offering structured workflows to manage email, calendar, contacts, teams, and meetings within the Spark ecosystem. It comprises a base use-spark skill for command reference, task-specific "recipes" (e.g., recipe-inbox-zero), and "personas" that define an agent's long-running role and workflow style (e.g., persona-exec-assistant), enabling LLMs to perform complex, context-aware operations.

Repo