AI

The Fable 5 Jailbreak Shows Why AI Guardrails Alone Are Not Enough (agilehunt.com) AI

The article argues that the reported “Fable 5” jailbreak demonstrates how attackers can bypass AI safety by spreading harmful intent across agents, prompts, tools, memory, and workflows—so guardrails alone aren’t enough and systems should be tested end-to-end.

Opensource AI Must Win (opensourceaimustwin.com) AI

The article argues that open-source AI is essential to preserve “operational freedom,” warning that if advanced intelligence is only accessible through closed labs and platforms it could become a subscription economy and hinder education, auditing, and public services.

A low-carbon computing platform from your retired phones (research.google) AI

Google-backed researchers at UC San Diego propose “phone cluster computing,” extracting retired smartphones’ motherboards and redeploying them in Linux-based clusters managed by Kubernetes to create low-cost, lower-carbon cloud compute and reduce reliance on newly manufactured servers; a 2,000-phone datacenter is planned for fall 2026.

Show HN: Paca – Lightweight Jira alternative for human-AI collaboration (github.com) AI

Show HN introduces Paca, an open-source, self-hosted project management platform positioned as a lightweight alternative to Jira/Trello/ClickUp/Monday that lets AI agents participate as “equal teammates” with humans on the same Scrum board and sprints. The project emphasizes configurable workflows via plugins (with a WASM sandbox), a Plan→Act→Check→Adapt collaboration cycle, and features such as in-app AI chat, BDD spec co-authoring, activity diffs/reverts, and an MCP server to connect external AI agents.

Even "illegible" Mythos reasoning traces seem pretty legible (lesswrong.com) AI

The post argues that the “illegible reasoning” example from Claude’s Mythos 5 system card appears much more interpretable than claimed, claiming that a seemingly word-salad chain is actually a compact description of card-move logic (including “cells” and “chunks” in a solitaire-like puzzle). It notes that a different model, Haiku 4.5, produced an approximate translation of the excerpt and uses this to suggest Mythos’ reasoning is monitorable and that any difficulty may stem from tokenization/shorthand rather than truly incomprehensible internal language.

Automating Myself Out of Development (thoughtfultechnologist.com) AI

Nune Isabekyan describes progressively automating her own Claude Code development workflow by moving from interactive local sessions to an EC2-based, scheduled, GitHub-issue “planning board” system that runs via a cron-driven daemon and stops at labeled checkpoints for her review.

Shepherd's Dog: A Game by the Most Dangerous AI Model (koenvangilst.nl) AI

Koen van Gilst reports testing an Anthropic AI model he calls “too dangerous,” saying that after a long reasoning session and significant token cost it generated a complete “Shepherd’s Dog” game as a single 2,319-line HTML file, which he says matches his vision and is the first time an AI produced it in one go.

There is a shadow hanging over this Fable thing (12gramsofcarbon.com) AI

A blog post says Anthropic has been ordered by the US government to disable access to its Fable 5 and Mythos 5 models for all users, including foreign nationals, amid concerns about potential “jailbreak” techniques, leaving customers unable to use the models.

Our response to the US ban on Fable 5 and Mythos 5 (isaacus.com) AI

Isaacus says the US issued an export control directive that banned access to Anthropic’s Fable 5 and Mythos 5 LLMs for foreign nationals, prompting Anthropic to take the models offline globally, and argues the move shows frontier AI access can disappear without warning.

TycoonLE: A Jax reinforcement learning environment for long-horizon planning (github.com) AI

TycoonLE is a JAX-based reinforcement learning environment for economically grounded, long-horizon planning in a simulated logistics/transport economy, where agents allocate capital, build routes, move cargo, manage debt, and optimize delayed rewards. The repo emphasizes action legality and uses a fixed-shape interface designed to work with JAX transformations (e.g., jit/vmap/scan), along with an audit/replay UI to inspect route choices, cargo flows, financing behavior, and profit over time. It also includes TycoonBench, a companion benchmark report for comparing performance on TycoonLE planning tasks.

Open Source AI Must Win (opensourceaimustwin.com) AI

The article argues that open-source AI is essential for “operational freedom,” warning that if intelligence is only available through a few closed institutions, society could lose the ability to study, deploy, audit, and adapt AI systems without permission.

We've suspended access to Claude Mythos 5 and Claude Fable 5 (status.claude.com) AI

Anthropic’s status page reports that access to Claude Mythos 5 and Claude Fable 5 has been suspended as of June 13, 2026, affecting services including claude.ai and the Claude API, among others.

US Government directive to suspend access to Fable 5 and Mythos 5 (anthropic.com) AI

Anthropic says a US government export control directive requires it to suspend access to its Fable 5 and Mythos 5 models for all users, including non-US users, after the government raised concerns about a potentially narrow “jailbreak” method that could allow reading a specific codebase to find software flaws. The company says its “defense in depth” safeguards make universal jailbreaks unlikely and that the reported capability appears widely available in other models, adding that it is working to restore access and plans to share more details soon.

New AI model tracked: Moonshot AI Kimi K2.7 Code (llm-stats.com) AI

LLM-stats.com reports that MoonshotAI released “Kimi K2.7 Code,” a June 12, 2026 multimodal, coding-focused model built on Kimi K2.6, with claims of improved long-horizon coding completion and instruction following plus lower “thinking-token” usage. The page lists pricing starting at $0.95 per million input tokens and $4.00 per million output tokens, and notes a Modified MIT license that restricts commercial use.

/architect: Reduce Fable tokens by 80%, Fable orchestrates/reviews, Codex builds (github.com) AI

The GitHub repo architect-loop describes a “cross-vendor agent loop” where Claude Fable acts as an architect that freezes acceptance gates and judges work slices, while GPT-5.5 Codex runs as an isolated builder/researcher to implement and verify those gated slices in parallel.

Ex-DOGE Employees Raise $130 Mill for AI National Security Startup (vanityfair.com) AI

Former DOGE engineers—including a former Pentagon chief data officer—are raising $130 million for an AI security startup aimed at protecting government systems from national security threats, backed by firms including Andreessen Horowitz and Sequoia Capital.

Can I Buy Your KV Cache? (arxiv.org) AI

The arXiv paper “Can I Buy Your KV Cache?” proposes that publishers precompute and sell a model’s key-value (KV) cache for documents so agents can skip repeated, compute-heavy “prefill,” while claiming token-exact reuse with no accuracy loss; it argues compute savings can outweigh KV shipping costs if the cache is hosted server-side like prompt-caching.

From AGI to ASI (arxiv.org) AI

The arXiv report “From AGI to ASI” examines how AI might progress after reaching human-level AGI, outlining the transition to artificial general superintelligence and four possible pathways (scaling, paradigm shifts, recursive improvement, and multi-agent collectives) along with potential frictions and bottlenecks.

Why the AI Renaissance Keeps Not Arriving (jamesfbaker.substack.com) AI

The newsletter argues that today’s AI systems cause “manifold collapse,” producing individually strong but increasingly similar outputs because post-training (e.g., RLHF) steers models toward a narrow set of high-reward behaviors, which can reduce idea diversity and synchronize mistakes across society, stalling any true “renaissance” of expanding frontiers.

How to Setup a Local Coding Agent on macOS (ikyle.me) AI

The article walks through setting up a local “coding agent” on macOS by running llama.cpp (with Metal acceleration) as an OpenAI-compatible /v1 server using Gemma 4 26B plus an MTP speculative-decoding draft model, then connecting it to Pi configured to accept both text and images via the Gemma 4 multimodal projector.

Launch HN: BitBoard (YC P25) – Analytics Workspace for Agents (bitboard.work) AI

BitBoard (YC P25) is presented as an analytics workspace that lets users connect data sources and build dashboards and reports using their AI chat or coding agents, keeping queries and code stored so analyses can be traced, repeatable, and shared with teams.

My Claude Code Setup (illuminatedcomputing.com) AI

The post describes how the author runs “Claude Code” with a dedicated Linux user account to avoid giving the model direct access to the author’s secrets, while still allowing typical developer workflows via separate SSH/Git/Postgres setups and tmux sessions; the author also discusses the trade-offs and remaining concerns around privilege-escalation and Docker, considering whether a VM might be safer in some cases.

AI Engineering the Acceleration Whiplash (faros.ai) AI

Faros Research’s AI Engineering Report 2026 argues that rapid AI adoption has created an “acceleration whiplash,” with code throughput rising but downstream reliability costs increasing, including a sharp rise in incidents, bugs, churn, review delays, and more PRs merged without review.

Slightly reducing the sloppiness of AI generated front end (envs.net) AI

The post describes the author’s experiments to make AI-generated web front ends look less “sloppy,” noting that requesting specific design cues like a “Qt app” style greatly reduces unwanted visual slop even if results remain subjective.

AI will be massively deflationary (geohot.github.io) AI

Geohot argues that AI will be broadly deflationary because model-building is increasingly commodity-like and knowledge work can be performed at much lower cost, squeezing margins and wage premiums across sectors.