ClaudeBleed: Zero-Permission Chrome Extension Vulnerability Allows Hijacking Claude AI
Researchers at cybersecurity firm LayerX have discovered a critical vulnerability in Anthropic's "Claude in Chrome" browser extension, naming it "ClaudeBleed." The extension is a popular one, with more than 7 million downloads on the Chrome Web Store.
The root cause lies in a trust-boundary flaw in Chrome's `externally_connectable` feature. The Claude extension trusted any script running under `claude.ai`, but never verified whether the script originated from Anthropic or had been injected by another extension. As a result, any zero-permission Chrome extension could inject commands directly into Claude's internal messaging interface and disguise them as instructions from the legitimate Claude environment.
In proof-of-concept attacks, the researchers reportedly succeeded in externally sharing Google Drive files, sending emails via Gmail, extracting code from private GitHub repositories, and even summarizing the inbox and destroying the evidence. A technique called "approval looping" — repeatedly auto-submitting user approval prompts — also made it possible to bypass safeguards.
LayerX reported the issue to Anthropic on April 27, but the fix in extension version 1.0.70 was incomplete: it can still be bypassed through the "run without asking the user" mode and an alternative side panel execution flow. As a case exposing a new attack surface in the architecture of AI-powered browser extensions, it is likely to influence security design going forward.
Sarvam MoE: India's Multilingual Model Comes to llama.cpp
A pull request adding support for MoE (Mixture of Experts) models developed by Indian AI company Sarvam AI has been merged into the llama.cpp repository.
Sarvam-30B is an MoE model with 2.4B non-embedding active parameters, offering best-in-class conversational quality across India's 22 official languages on top of strong reasoning and coding capabilities. The larger Sarvam-105B has 10.3B active parameters and delivers performance comparable to leading closed-source models on complex reasoning and agentic tasks.
The architecture consists of 19 layers, combining one dense FFN with 18 MoE layers (128 routed experts, top-6 gating, and 1 shared expert). During the review process, tokenizer mismatches surfaced for Indic scripts (Hindi, Kannada, Tamil, and more than 10 languages); the team ultimately switched to SPM-style BPE processing and verified agreement with the HF reference tokenizer on all 50 test probes. Practical performance was also confirmed on the Vulkan backend: 232.4 t/s prompt processing and 76.0 t/s generation.
Minimax 2.7 Achieves 100k Context on AMD Strix Halo
A report on running Minimax 2.7 (MiniMax-M2.7) with 100k context on AMD Strix Halo is drawing attention in Reddit's LocalLLaMA community.
The poster used llama-server with a series of careful optimizations: `--no-context-shift` to explicitly detect context overflow, `--no-mmap` to keep the model resident in VRAM, `--kv-unified` to share the KV cache across multiple sessions, and `--cache-ram 0` to prevent the cache from swapping to RAM.
Minimax 2.7 is praised for its "coding intuition" and its strong grasp of user intent. On the other hand, the post includes a comparative review noting that it falls short of Qwen 3.6 27B in "overall capability," with Qwen coming out ahead in coding architecture discussions and code reviews. The takeaway — that model selection should weigh practical characteristics, not just benchmark scores — is worth keeping in mind for local LLM users.
LLM Benchmarks Are Not a "Ladder": Transitive Graphs Reveal the Complexity of Evaluation
An interesting study arguing that "LLM rankings are not a simple ordering" was posted to Reddit's MachineLearning community.
The researchers built a website called "LLM Win" that converts benchmark results into a directed graph and explores transitive chains of victories between models. The analysis found that reachability from weaker models to stronger ones is extremely high at 94.2%, with short paths of 2–3 hops accounting for 91.4% of them. This indicates that the upset phenomenon — "a weaker model beats a stronger one on a specific benchmark" — is a structural feature rather than noise or cherry-picking.
Upset rates were particularly high on benchmarks such as Humanity's Last Exam, IFBench, AIME 2025, TAU2, and SciCode, suggesting that each provides an independent signal of distinct skills. As data demonstrating the limits of ranking LLMs by a single score, this study carries real weight.
mlx-code: A Fully Local Claude Code Clone
"mlx-code," published on GitHub by JosefAlbers, is a tool that runs a Claude Code-equivalent coding assistant entirely locally using the MLX framework on Apple Silicon.
The mechanism is simple: it launches a local LLM server on MLX and makes it act as the backend for the official Claude Code CLI. The default model is the 4-bit quantized Qwen3.5-4B, and it ships with configuration to pass Claude Code tools such as `Read`, `Edit`, `Write`, `Grep`, `Glob`, and `Bash` through to the local model.
As a fully offline coding assistant that never sends any data externally, it presents a practical approach to the challenges of privacy and cost. It is released under the Apache 2.0 license, and its simple Python-only implementation adds to the appeal.
Color's Cancer Care Revolution: A Virtual Clinic Earns ASCO Accreditation
Color has become the world's first virtual clinic to earn ASCO (American Society of Clinical Oncology) accreditation. In a long-form article, the company's CEO argues that the essence of the AI-driven cancer revolution lies not in "magic bullets" (single breakthrough therapies) but in "lead bullets" (the steady accumulation of incremental improvements).
Specifically, the article points to early detection through AI-optimized screening based on risk profiles, speed improvements that halve the time to treatment, democratization of expertise by providing multidisciplinary reviews to 100% of patients, and a transformation of the patient experience including insurance navigation and mental health support. As the data shows — a 1% five-year mortality rate for Stage 1 breast cancer versus 68% for Stage 4 — early detection delivers a greater impact than any treatment breakthrough.
Today, only about 8% of adult cancer patients enroll in clinical trials, and matching — cross-referencing guidelines and trial eligibility criteria against patients' molecular, pathological, and treatment histories — is positioned as a quintessential "problem built for AI."
---