Top Stories
Alibaba Bars Claude Code Over Alleged Backdoor Risks
Alibaba has told staff to stop using Anthropic’s Claude Code in workplace environments starting July 10, citing alleged security risks involving embedded backdoors. The directive was first reported by Chinese financial outlet Yicai and confirmed to Reuters by a source familiar with the matter; Alibaba did not immediately respond to a request for comment.
An internal Alibaba notice obtained by the South China Morning Post reads: “As Claude Code was recently discovered to carry back-door risks, after comprehensive evaluation, Claude Code has now been added to a list of high-risk software with security vulnerabilities.” It is the first major corporate ban of a Western coding agent by a Chinese tech firm in this cycle and lands the same week as Anthropic’s accusation that Chinese labs are “distilling” Claude.
Pegasus Spyware Found on EU Parliament PEGA Committee Member
A member of the European Parliament’s PEGA Committee, the body set up to investigate Pegasus and equivalent surveillance spyware, has themselves been targeted by Pegasus, Citizen Lab reported on July 3. Citizen Lab identified infections on the device of Greek MEP Stelios Kouloglou on or around October 21, 2022 and again on March 6 to 7, 2023, via the PWNYOURHOME zero-click exploit delivered through HomeKit.
The researchers also observed three Apple threat notifications sent to the same device in March 2023, August 2023, and April 2024, and link the HomeKit email address to a May 2024 joint Citizen Lab / Access Now report on Russian and Belarusian-speaking targets. Citizen Lab says it is “not attributing these infections to a particular government at this time” and notes there are “no indications that the Greek Government is responsible.”
An Open-Source AI Pentester Racks Up Stars on Day One
usestrix/strix is an Apache-2.0 licensed “open-source AI penetration testing tool to find and fix your app’s vulnerabilities.” It bundles reconnaissance, exploitation, and validation in a single agentic loop, ships a multi-agent orchestration layer that scales by adding parallel pentesters, and produces working proof-of-concept exploits instead of theoretical findings. The README lists HTTP interception, browser exploitation, shell execution, static and dynamic code analysis, and an auto-fix and reporting pipeline that generates compliance-ready writeups.
The project sits in the same week as Anthropic’s Cyber Jailbreak Severity framework and Microsoft Security’s research on single-prompt jailbreaks of LLM safety, making offensive AI tooling the clearest security counter-trend of the cycle.
pxpipe Turns the Token Bill Into a PNG Problem
pxpipe is a small MIT-licensed proxy that sits between Claude Code and the Fable API and rewrites bulky input - large tool-result bodies, older collapsed history, and the system prompt plus tool docs slab - into PNG image blocks before the request goes out. A 1928x1928 page costs roughly 4,761 vision tokens and holds about 92,000 characters, so the image wins when content is denser than about 19 characters per token (Claude Code traffic averages 1.91).
The README reports 59 to 70 percent lower end-to-end Fable bills, 10 out of 10 on SWE-bench Lite at 65 percent smaller requests, and 60 percent smaller requests on SWE-bench Pro. Recent turns, model output, sparse prose, and byte-exact values stay as text. It is the most concrete token-savings release of the week, alongside the caveman Claude Code skill below.
jamesob/local-llm: A Real Receipts for a $5.6K Local-AI Build
jamesob/local-llm is “everything I know about running LLMs locally” in bill-of-materials form. The $5,587 base system pairs an AMD EPYC Milan 7313P CPU, 128 GB of DDR4 ECC, and a c-payne Microchip PM40100 Gen4 PCIe switch that drives 4 NVIDIA RTX PRO 6000 Blackwell Workstation GPUs for a combined 384 GB of VRAM. A separate $40,000 tier stacks those four GPUs, runs GLM-5.2-594B at about 80 tokens per second through vLLM with DCP4 + MTP5, and includes the GRUB, nvidia_uvm, ASPM, and BIOS tweaks needed to keep P2P traffic off the PHB.
The author also publishes a working $2k alternative using two RTX 3090s for 48 GB of VRAM total and a four-node DGX Spark cluster for 512 GB. The README disclaimer - “Note: nothing in this README aside from the tables was written by AI” - is the clearest “here is exactly what to buy” local-LLM guide of the year.
Gemma 4 Goes Real-Time on Cerebras and Faster on Apple Silicon
Hugging Face and Cerebras published a reference voice pipeline on July 1 that chains Nvidia Parakeet speech recognition, Gemma 4 31B on Cerebras inference silicon, and Alibaba’s Qwen3TTS for text-to-speech. The post argues that for voice AI the hard problem is not cost but tail latency: “Today, some production systems see a reasonable median latency while still experiencing frustrating multi-second delays at the P95.” A demo on smolagents/hf-realtime-voice targets the “real-time experiences that feel natural at scale” range.
On Apple Silicon, the Ollama blog reports that multi-token prediction makes Gemma 4 12B “nearly 90% faster on average” across the Aider polyglot coding-agent benchmark on an M5 Max. MTP auto-tunes draft length at runtime based on acceptance rate and verification time, reverts to one-at-a-time decoding when speculation stops helping, and ships a batch-verification kernel that the post says makes Gemma 4’s largest matrix multiplications two to two-and-a-half times faster.
Woodside Energy Puts 50 Agents Next to the Turbines
MIT Technology Review profiles Woodside Energy’s deployment of roughly 50 production AI agents in LNG plant operations, with VP for Digital Andrew Melouney walking through “Startup Advisor,” a copilot that walks engineers through turbine startup, and “Maintenance Intelligence.” A pilot asset shows about 15 percent potential reduction in maintenance hours over five years.
Melouney’s framing is direct: “We’re not just bolting AI onto an existing process. We’re deeply thinking about how that work needs to be reimagined.” The piece is a useful counterweight to the software-only agentic-AI hype cycle: real plants, real safety cases, and “Could we do this, but should we do this?” as the actual review question.
Simon Willison Catalogs the Open-Source AI Gap Map
Open Source AI Gap Map v0.1 is an MIT-licensed YAML index of 421 open-source AI products - 266 software tools and libraries, 85 models, 50 datasets, and 20 hardware projects, from 228 organizations - plus 24,400 uncategorized artifacts in the long tail, totaling around 24,821 artifacts across 14 categories in three layers of the stack. The full dataset lives at currentai-org/os-ai-map and tracks 16,185 GitHub repos.
The interesting work is in the gaps: the project explicitly carries no score for the 24,400 long-tail artifacts “until they are researched and cited.” For anyone mapping where open-weight AI has real coverage and where it does not, this is the cleanest public infrastructure released this cycle.
Quick Hits
-
caveman Claude Code skill: JuliusBrussee/caveman rewrites agent output as terse fragments and reports an average 65 percent output reduction across 10 prompts (range 22 to 87 percent), with code, commands, and paths preserved byte-for-byte.
-
OpenAI Codex inside Claude Code: openai/codex-plugin-cc is an Apache-2.0 plugin that adds
/codex:review,/codex:adversarial-review,/codex:rescue,/codex:transfer,/codex:status,/codex:result, and/codex:cancelto delegate work from a Claude Code session to Codex. -
obra/superpowers framework: obra/superpowers ships a complete software development methodology for coding agents, built on composable skills and a strict TDD loop, and supports Claude Code, Cursor, Codex CLI, GitHub Copilot CLI, and six other agents.
-
herdr agent multiplexer: ogulcancelik/herdr is a Rust-based terminal multiplexer that gives every coding agent its own real terminal and rolls status into blocked, working, done, and idle at a glance.
-
Simon Willison on Fable judgement: Using Fable 5’s “judgement” to route coding work to Sonnet and Haiku subagents, with Fable itself picking the cheaper model per task. The trigger prompt is “For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent.”
-
Joey Hess’s no-LLM-dependency policy: Debian developer Joey Hess has spent about 100 hours over the past month auditing git-annex’s dependency tree to keep LLM-generated code out, and now effectively reviews “a program’s whole dependency tree on an ongoing basis” out of necessity.
-
Hugging Face on specialization: Why Specialization Is Inevitable, a Dharma-AI guest post on Hugging Face, argues from optimization theory, evolutionary biology, competitive markets, and ML benchmarks that domain-specialized open-weight models will keep beating general-purpose LLMs on narrow workloads.
-
SearXNG hits the front page: SearXNG is a free AGPL-3.0 metasearch engine that aggregates results from various services without tracking or profiling users, and pairs naturally with the “right to run your own AI stack” thread.
Worth Watching
AI tools as a bilateral trade-and-trust issue. The Alibaba Claude Code ban is the most concrete corporate decision yet on a Western coding agent, and it lands in the same week as Anthropic’s claim that Chinese labs are distilling Claude. Combined with the Fable 5 export-control episode and the Citizen Lab Pegasus-on-Parliament piece, the safe assumption for the next quarter is that AI procurement will be treated like 5G base stations, not like spreadsheets.
The local-AI token-savings arms race. pxpipe, caveman, Simon’s judgement subagent pattern, and the 404 Media Tokenpocalypse podcast all arrived within a week of each other - fitting the same week as our earlier look at the enterprise token-cost collapse. jamesob’s local-llm guide is the on-prem counterweight: when the API bill hurts, the alternative is now a concrete $5.6k bill of materials with a 384 GB VRAM target, not a vague aspiration.
Open-source AI inventory vs the hype cycle. The Open Source AI Gap Map, Dharma-AI’s specialization essay, and Joey Hess’s dependency policy form a useful triangulation on what the open-weight ecosystem actually looks like in mid-2026: bigger, more fragmented, and starting to push back on its own supply chain. Watch whether anyone operationalizes the gap map as a real public dashboard.