AI Daily — 2026-07-02

English 中文

Mythos AI era shows AI detects vulnerabilities at scale · PerceptionRubrics Calibrates Multimodal...

Covering 38 AI news items

🔥 Top Stories

1. Mythos AI era shows AI detects vulnerabilities at scale

Epoch AI notes that Mythos-enabled systems, including Fable, are finding software vulnerabilities at scale. In June 2026, 21 notable organizations disclosed roughly 1,500 high- to critical-severity CVEs, over 3.5x the monthly record prior to Claude Mythos Preview. Source-twitter

2. PerceptionRubrics Calibrates Multimodal Evaluation to Human Perception

PerceptionRubrics introduces a rubric-based evaluation framework designed to narrow the gap between benchmark scores and real-world brittleness in multimodal models. It pairs 1,038 information-dense images with over 12,000 instance-specific rubrics, derived from golden captions using a Circular Peer-Review consensus pipeline and distilled into a dual-stream evaluation system. Source-huggingface

3. ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving

A new decode router, ELDR, targets PD-disaggregated MoE serving by considering expert locality rather than just load. It uses prefill expert activations to route requests to decode workers whose loaded experts match the batch, aiming to reduce latency variability in MoE models. The approach addresses a key inefficiency in existing routers for MoE deployments. Source-huggingface

📰 Featured

LLM

MOTHRAG: Graph-Free Multi-Hop Retrieval Open-Sourced — MOTHRAG, a graph-free multi-hop RAG framework, has been open-sourced to avoid offline knowledge graphs and costly re-indexing. It uses a dense index with query-time orchestration (no graph, no GPU) and supports updates via embed-and-append without full rebuilds. Benchmarks on HotpotQA, 2WikiMultiHopQA, and MuSiQue show competitive accuracy and F1 compared with graph-based systems like GraphRAG, HippoRAG, and RAPTOR. Source-reddit
GPT-5.6 Sol Ultra Teased for Codex Rollout — A social post hype-promotes GPT-5.6 Sol Ultra, hinting at a ‘power of the sun’ coming to Codex. The poster, Tibo, expresses eagerness for how people will use it and advises saving hard prompts, signaling buzz around a new AI model. Source-twitter
Anthropic Guardrails Drive Fable 5 Rerouting to Opus — A benchmark discussion indicates that many triggers cause rerouting of Fable 5 to Opus, driven by guardrails set by Anthropic. The upshot is that Fable 5’s performance remains the same, but the probability of rerouting has increased. Some users push back, arguing the problem lies in the hard guardrails rather than the model itself. Source-twitter
SkillHone Enables Continual Agent Skill Evolution via Persistent History — SkillHone introduces a harness for continual agent skill evolution grounded in persistent decision history, enabling future agents to interpret prior revisions, evaluations, and rejected alternatives. It addresses the limitation of existing methods that only retain the final artifact from bounded runs and discard the decision history. By preserving this history, SkillHone aims to improve task-specific procedures, scripts, and references across changing tasks and environments. Source-huggingface
AllenAI olmocr v0.4.0 Enables PDF-to-Text OCR for LLMs — olmocr is an open-source toolkit that converts PDFs and image-based documents into clean, plain text for LLM datasets, with an online demo available. It converts PDFs, PNGs, and JPEGs to Markdown, preserving equations, tables, handwriting, and complex formatting, while stripping headers and footers and preserving a natural reading order. The GPU-accelerated process is advertised under $200 per million pages, with v0.4.0 adding RL training and a ~4-point boost on olmOCR-bench. Source-github

AI Tools

Claude Code Artifacts expand to Pro and Max plans — Claude Code’s Artifacts are now available on Pro and Max plans, enabling live, interactive pages built from your session. Claude writes the code, publishes it live to claude.ai, and updates it in real time while it keeps working. Pages are private to your account, fully self-contained, and shareable as private dashboards or PR-like demos; beta access is also available on Team and Enterprise plans. Source-twitter
Fable xhigh with GPT-5.5 xhigh as planner, coder, judge — A user reports success using Fable xhigh as a planner/architect, GPT-5.5 xhigh (subscription) as a coder, and Fable xhigh again as a judge. They note that planning and judging cost only a few dollars, versus typical $50+ full round trips, and praise GPT-5.5 for being cheap and fast, though the approach’s longevity remains unclear since the re-release is recent. Source-twitter

RL

Introducing EBR-bench: measure on-the-fly learning with Earthborne Rangers — Epoch AI Research introduces EBR-bench, a benchmark to measure on-the-fly learning by having AI repeatedly play Earthborne Rangers. Early results show no signs of improvement, underscoring the difficulty of rapid adaptation. The benchmark aims to quantify learning from mistakes in dynamic game settings. Source-twitter

LLMs

Anthropic’s Claude Design sparks partner trust concerns with Figma — According to The Information, Anthropic blindsided its then-business partner Figma with the launch of Claude Design, and Figma’s founder said Anthropic had not been consistently honest; Anthropic’s chief product officer had served on Figma’s board until three days before the launch. The move coincided with a drop in Figma’s stock and a surge in Anthropic’s valuation, fueling debate about competition, data control, and partner dynamics. Dario Amodei argues that open-source models powerful enough to rival Anthropic could be dangerous, underscoring tensions around moral branding, partner access, and rapid vertical expansion. Source-twitter

Open Source

Karukan: Open-Source Neural Kana-Kanji IME — Karukan is an open-source Japanese input method for Linux and macOS featuring a neural Kana-Kanji engine. It performs inference with a GPT-2–based model via llama.cpp, enables live, context-aware conversion with on-device learning, and ships with a system dictionary derived from SudachiDict, a candidate rewriter ported from Mozc, emoji input, and Slack-style triggers. The project is modular, offering Linux/frontends (fcitx5), macOS frontend, a shared engine, and CLI/server tools; the first launch downloads the model from Hugging Face, causing initial delay. Source-github
SentryCode Real-Time Auditor and Honeytokens for AI Coding Agents — Open-sourced SentryCode is a kernel-level auditing tool designed to monitor local AI coding agents. It logs file, network and cue activity, uses honeypot tokens for zero-false-positive data breach detection, detects covert channels, and provides tamper-proof audit logs plus policy enforcement, all running locally with no outbound connections. A demo binary is available on GitHub and feedback from local AI-agent users is welcomed. Source-reddit

AI Infrastructure

CubeSandbox Delivers Fast, Secure AI Sandbox for Agents — CubeSandbox is a high-performance sandbox service for AI agents built on RustVMM and KVM. It supports single-node and multi-node deployments, delivering hardware-isolated sandboxes with startup under 60ms and memory overhead under 5MB, while remaining compatible with the E2B SDK. The v0.4 release adds safer egress, a credential vault, a dashboard with health checks, and the CubeCoW Copy-on-Write snapshot engine introduced in v0.3.0 for event-level snapshots and rapid snapshot operations. Source-github

AI Safety

Gnosys Improves Safety Classifier with Sparse Labels — Gnosys Labs demonstrates an autonomous model engineering approach that improves prompts and classifiers when ground-truth labels are scarce. On the ToxicChat safety benchmark, it outperformed the starting classifier and GEPA under a fixed 5% false-positive rate across two held-out runs, with results and limitations discussed in the note. Source-reddit

Multimodal

MARS2 Workshop at ECCV 2026: Multimodal Video Reasoning — The MARS2 Workshop (Multimodal Reasoning Competition) at ECCV 2026 targets multimodal and test-time (slow) reasoning, with emphasis on video and real-world tasks like advertising understanding. It gathers researchers from MIT, Cambridge, Oxford, CMU, NTU, and lists Tec-Do and Minimax as organizers/sponsors. The post questions the practical evaluation setup and the usefulness of this benchmark for day-to-day development. Source-reddit

⚡ Quick Bites

Updated AI Dictionary Satirically Defines Common Jargon Terms — An AI-focused Twitter thread presents an updated AI dictionary that humorously redefines familiar jargon. The list covers terms such as The bitter lesson, AGI, RL, continual learning, and transformer-related ideas with satirical explanations. It lampshades AI research culture by poking fun at how terminology is used. Source-twitter
TurboServe Enables Efficient, Scalable Streaming Video Generation — Streaming video generation is described as a new serving workload with long-lived sessions that produce video chunk by chunk. It requires preserving session state across active and idle periods and delivering chunks within tight latency targets. In multi-user, multi-GPU environments, two key challenges are session duration heterogeneity and the need to repeatedly schedule ongoing sessions. Source-huggingface
GitHub dataset: multilingual fitness exercises metadata and backend scaffold — The repository hasaneyldrm/exercises-dataset provides a structured dataset of fitness exercises with metadata such as name, category, target muscle group, equipment, and instructions, plus media references. It also offers a developer setup wizard to scaffold a back-end for 1,324 exercises in six languages (English, Spanish, Italian, Turkish, Russian, Chinese), while noting that exercise media files are not included due to ownership claims. Source-github
How Best Paper and Oral/Highlight Selections Work at ML/CV Conferences — The post questions how papers are chosen for Best Paper, Oral, or Highlight at top ML/CV conferences. It asks who selects candidates (ACs, SACs, program chairs, award committees, or a separate group), whether camera-ready versions are used, and whether decisions rely on reviewer scores or factors like novelty and discussion. Source-reddit
Map of 11M papers by semantic similarity and time slices — Researchers have built a free visual map, The Global Research Space, that aggregates the latest 11 million papers from OpenAlex and arXiv. Papers are encoded with SPECTER 2 on titles and abstracts, reduced to 2D with UMAP, and labeled within Voronoi regions around high-density peaks. The tool supports keyword and semantic search, offers analytics on institutions, authors, and topics, and includes a time slider with daily auto-ingestion for up-to-date coverage. Source-reddit
System-Level Prompt Injection Guard Separates Instruction and Data Channels — Proposes a middleware-based mitigation for prompt injection in tool-using LLMs, introducing the Sentinel Gateway to strictly separate the instruction (trusted) and data (untrusted) channels. The system enforces signed, scoped runtime authorization tokens for all agent actions, decoupling observation from execution. Implemented as a FastAPI middleware with token-based execution, plus a Streamlit interface for inspection and audit logs. Source-reddit
Improving a 5-Class Diabetic Retinopathy Model (APTOS 2019) — A final-year student builds a Flask-based AI Diabetic Retinopathy detection app using a 5-class classifier trained on APTOS 2019. The model shows inconsistent predictions across classes, misclassifying some cases and sometimes producing high confidence on incorrect results, especially with non-APTOS images. Efforts to improve include trying pretrained models such as ResNet50 (trained on APTOS) and ResNet152. Source-reddit
Interview Tomorrow with Hermes Co-founder karan4d — An interview is planned with Hermes co-founder karan4d of NousResearch. The host invites audience topics, noting prior coverage on setup, integrations, and basic cron jobs, and asking what else to discuss. A lighthearted remark about a chart being a vertical line is included. Source-twitter
Hiring: Strong Product Manager for AI Inference, Urgently — A Twitter post announces an urgent need for a highly capable product manager focused on AI inference. The call emphasizes immediacy in recruiting for this AI-inference leadership role. This reflects ongoing demand for product leadership in AI inference initiatives. Source-twitter
Open Source Spurs Unlimited Sovereign Intelligence — The tweet proclaims that open source equals unlimited sovereign intelligence. It suggests that open-source AI enables sovereign capabilities for users or nations. The post reflects optimism about open-source AI’s democratizing potential. Source-twitter
Books and resources to strengthen ML foundations — An ML PhD student seeks recommendations to bolster mathematical foundations, focusing on Linear Algebra, Probability Theory, and Functional Analysis. They mention Linear Algebra Done Right, A Primer on RKHS, and potentially re-reading PRML, with plans to explore Pat Kidger’s work as well. The goal is a thorough fundamentals refresh before graduation. Source-reddit
Differential-Geometry View on Hamiltonian Neural Networks — A write-up exploring Hamiltonian Neural Networks (HNN) from a differential-geometry perspective rather than the standard loss-centric framing. The author highlights links to Noether’s theorem, emphasizing symmetries and conservation laws in physics-informed neural networks, and presents math-heavy insights with visuals and explanations. Source-reddit
Improving Machine-Translated Webnovels via Style Transfer: Faithfulness vs Fluency — Researchers discuss using style transfer to clean up machine-translated webnovels, aiming to produce prose that reads like professional writing while staying faithful to the original text. They consider STRAP-based unsupervised approaches, including paraphrase framing, pseudo-parallel data generation, and a style-specific inverse model, due to lack of clean parallel data. A major hurdle is maintaining long-range context across thousands of pages. Source-reddit
Has Anyone Tried Fast Byte Latent Transformers with Mamba? — Reddit user references the paper arXiv:2412.09871v1 and asks whether replacing the entropy-model transformer with a Mamba model could reduce computation (O(n)). The post invites discussion on feasibility and potential changes in approach. Source-reddit
BMVC 2026 Review Thread Opens Tomorrow — The post announces that BMVC 2026 reviews will be released tomorrow and creates a dedicated parent discussion thread for the community. It mentions the author and provides a link to the Reddit discussion. Source-reddit
Describing Efficient Models: Higher Accuracy, Fewer Params — An online poster asks for guidance on how to describe a model that achieves higher accuracy with fewer parameters and FLOPs. The discussion touches on benchmarking, model compression, and the rhetoric used to communicate efficiency gains. It’s a request for clarification from a novice in the AI community. Source-reddit
Anthropic model praised amid Fable critiques — An X user identifies as Anthropic’s #1 hater but says the model is good. They call ‘Fable’ takes dumb and argue critics distract from real issues Anthropic should address. The post frames support for Anthropic’s work while criticizing opponents. Source-twitter
Govt Stakes in AI Firms Must Be Voluntary, Not Tax-Based — The post discusses rumors that the government could take stakes in AI companies, insisting two conditions: contributions must be voluntary rather than a forced tax, and funds should go to citizen-owned accounts rather than a government slush fund. It stresses that the government would coordinate the effort but not own it outright. Source-twitter
T3 Community Meet-and-Greet at OpenAI Booth Today — The T3 Community is hosting a meet-and-greet today at 2pm at the AI Engineer venue, with plans to be at the OpenAI booth. Attendees are invited to visit the OpenAI stand for networking. Source-twitter
Fan Expresses Love for Fable 5 and Anthropic — On Twitter, a user expresses affection for Fable 5 and Anthropic. The post blends enthusiasm for a video game with admiration for an AI safety company, illustrating cross-domain interests in a single social post. Source-twitter
ACL ARR May 2026 Reviews: When Are Results Out? — A Reddit user asks whether ACL ARR May 2026 review results will be released on July 2 or July 7, and what the acceptance thresholds are for the Main and Findings tracks. The post indicates the author is new to ACL submission outcomes and seeks guidance. Source-reddit

Generated by AI News Agent | 2026-07-02