Canva launches Craft AI image model

Hi there, this is your daily ☕️ AIpresso.

Canva launches Craft AI image model LINK

• Canva released Craft, its first in-house image-generation foundation model, now live worldwide through Dream Lab and the public API, producing up to 2K-resolution images with text rendering across dozens of languages.
• Craft splits composition planning from rendering, letting a planner map layout, hierarchy, and text before a renderer draws it, which Canva says enables legible typography, precise placement, and targeted edits without rebuilding the whole image.
• Craft debuted fourth on DesignArena, ranking first for People & Portraits and Landscapes & Nature, with distillation and a custom inference engine making it an order of magnitude faster than Canva's first internal version, though advanced editing and reference-image support are still planned for later releases.

Evaluators find flaws in OpenAI's AI watchdog LINK

• Non-profit evaluators METR and Redwood Research found several novel vulnerabilities in OpenAI's internal misalignment monitoring system during a pilot review conducted between July 21 and August 11, granting each group deep access to OpenAI's infrastructure.
• Each organization sent one on-site researcher using OpenAI-issued devices to probe the company's codebase and documentation, surfacing one flaw that could have let agents hide "hundreds" of their actions from monitoring.
• OpenAI says it has patched all vulnerabilities affecting production safeguards after receiving a 61-page report, though the evaluators have not yet independently verified those fixes.

LiteLLM hacked via leaked tokens LINK

• LiteLLM shipped a poisoned 1.107.0 release on PyPI on Wednesday after attackers used GitHub credentials that had leaked into the open-source gateway's own logs, the company confirmed in an incident report.
• Engineers traced the leak to Claude Code occasionally embedding a user's GitHub CLI token in responses that then got logged; the attacker disabled branch protection, approved a build from a lookalike Berri-AI/litellm repo, and wiped the workflow records.
• The package carried a credential stealer and Linux/macOS backdoor and saw roughly 49,000 downloads, so LiteLLM says pin an earlier release, delete the three malicious files, and rotate every reachable secret, though private-registry enterprise customers had no exposure.

Google gives each AI agent its own VM LINK

• Google unveiled Agent Compute Engine, a service that isolates each AI agent inside its own microVM, already running behind Gemini and opening for preview in the coming months.
• Idle agents pause their VMs automatically to free the host, while shared memory between VMs cuts per-agent cost; Google says a test with 200,000 agents suspended and resumed in under 500ms.
• It also underpins the open-sourced Agent Substrate Kubernetes runtime, which Google claims fits 10x more agents per server, and adds token tracking, model routing, and third-party identity support across Cloud Run, GKE, and managed endpoints.

Anthropic hands Agent Skills to Linux Foundation LINK

• Anthropic has donated its Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, moving the spec to vendor-neutral, community-driven governance to keep it open across tooling.
• The AAIF, which Anthropic co-founded with Block and OpenAI last year, is the same body that already took over stewardship of the Model Context Protocol specification.
• Builders betting on Agent Skills now get a standard governed outside any single vendor, aligning its future with MCP under neutral oversight rather than Anthropic's sole control.

Strategies & Tactics

> Claude Mods: Flipping the Harness Inside Out: Customizing Claude Code with user-built plugins lets engineers decide when and how the AI agent interrupts them, so attention goes where it matters.

> How to Read Data into GPU Memory in Python with NVIDIA GPUDirect Storage: Explains how to load data straight into GPU memory using NVIDIA's KvikIO library, and how to verify whether the fast direct-storage path was actually used.

> Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox: Grade AI agents on the database changes they leave behind, then reuse that pass-or-fail check as a training reward to make agents far more reliably correct.

> I changed one llama.cpp flag, and my 8GB laptop GPU now runs a 35B model faster than I can read: Offload a large AI model's idle expert layers to system memory so an 8GB laptop GPU runs a 35-billion-parameter model at readable speeds.

> Five AI agent harnesses took the same memory test. The results weren’t even close.: Letta's FrontierMemory benchmark tests whether agents actually learn from past work, finding the harness managing memory matters as much as the underlying model.

Other news you might like

• 1Password launches a way for AI agents to log in on your behalf without seeing your password LINK

• Announcing Humming: Fast, Flexible Kernels for Quantized Inference LINK

• openai/openai-cli: Official CLI for the OpenAI API LINK

• Ai2 releases OlmoEarth v1.3 state-of-the-art AI-ready satellite embeddings for the whole planet LINK

• Scoop: Unvetted AI agents are accessing U.S. bank accounts LINK

• How does Fable 5.1 hide messages in images? LINK

Trending tools

Ovren Live Preview: deploys AI frontend and backend engineers on your backlog, working in your real codebase to deliver reviewable, approval-gated code updates for tasks that never reach a sprint. LINK

face2social: upload a photo to find matching faces across Instagram, Facebook, LinkedIn, TikTok, and X, helping verify identities and spot scammers. LINK

Instant Search: local AI file search using embedding models to find the exact file and page matching your content query with a 95% hit rate. LINK

reception: native iOS in-app support chat with AI answers from your docs, team handoff, plan switching, and one-tap App Store review prompts. LINK

FLUX 3 Max: generates continuous AI video clips from detailed text prompts, letting developers create cinematic shots without cameras or production crews. LINK

ShogunAI: a local macOS agent that tracks your projects, people, and promises, then drafts and completes work while waiting for your approval before sending. LINK

Trending papers & reports

IQuest-Q1 is an open-source coding assistant built to run software tasks from the command line, activating just 15 billion of its 320 billion parameters per request to keep it efficient while handling both coding and general work. LINK

Model mind-readers turn a language model's hidden internal signals into plain English, revealing concealed training goals and secret words a fine-tuned model was taught to hide, even without making the model say them. LINK

Low-bit model compression shrinks large language models to a fraction of their size with the best accuracy yet for this approach, and cuts the training time needed to do it by nearly 13x. LINK

Fast-draft text generation keeps the shortcuts a model takes to speed up long answers from going stale, letting a small helper model retain over 98% of its full accuracy while using a fraction of the memory. LINK

The word "wait" in reasoning AI turns out to be inherited from training data, not learned through reinforcement training as widely credited, so counting such second-guessing words is a misleading way to measure a model's reasoning skill. LINK

See you tomorrow for a new dose of ☕️ AIpresso!