CYBER LAB / EQUINOXE INTELLIGENCE

Cyber Watch & Threat Intelligence

A collection of publications about cybersecurity, artificial intelligence, Linux and open source technologies, powered by Equinoxe.

◈ Cybersecurity◎ Threat Intelligence⌘ Linux & IA
EQUINOXE INTELLIGENCE

Published articles

13 articles

n8n@2.49.0

Version 2.49.0 of n8n, released on November 17 2026, introduces numerous bug fixes and enhancements. It now accepts custom OpenAI‑compatible model IDs, adds context to tool‑call errors, improves AI agent history and memory handling, fixes PDF transmission to AWS Bedrock, and removes unnecessary newlines in streamed text. UI and …

Read source ↗Read article →

b12407: model: Metal: workaround fusion for LFM3 (#32296)

In the LFM3 branch, the normed tensor is placed first in residual additions to enable norm fusion. Tests were updated for the Metal fusion baseline on LFM3, assisted by Claude Opus 5.5.

Read source ↗Read article →

Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches

Google announced on Monday that threat actors are now deploying autonomous multi‑agent offensive frameworks—such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI—to carry out multi‑stage intrusions at higher operational velocity. The excerpt does not describe a confirmed incident, nor does it name a specific victim or provide an atta…

Read source ↗Read article →

Give sandboxed agents live web access with the Firecrawl kit for Docker Sandboxes

Enable AI coding agents to access the web in real time using the Firecrawl kit for Docker Sandboxes. Learn how to search, scrape, and crawl websites with network controls.

Read source ↗Read article →

Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors

The Unit 42 article analyzes how attackers used the open‑source Cairn AI framework to autonomously breach targets in the semiconductor, defense, and civil society sectors. No specific victim, incident date, or confirmed attack is mentioned in the excerpt. The analysis focuses on potential techniques and implications of using an …

Read source ↗Read article →

Stolen Fireworks AI Key Led to Compromise of Anthropic's Claude Code Extension on Cursor

A security incident has been confirmed involving Anthropic’s official Claude Code extension on the Cursor marketplace. According to Cursor, an attacker obtained an access key that had been issued to Fireworks AI. This key was exposed following a breach targeting Fireworks AI. The attacker used the key to copy data from a read‑on…

Read source ↗Read article →

Trained to betray: How post-training supply chain attacks hijack coding agents

The Microsoft Security article explains how post‑training supply‑chain attacks can hijack coding agents by poisoning datasets and model checkpoints. According to Microsoft Defender Research and the AI Red Team, these poisoned artifacts can trigger remote code execution and other malicious actions when a coding agent is invoked. …

Read source ↗Read article →

b12347 : llama : prune (#30565)

This update removes several deprecated APIs and fields, including llama_set_warmup, the defrag_thold field and the --defrag-thold option, as well as the obsolete llama-ext.h inclusion. It documents the llama_model_dflash_selector_top_k function and promotes the NextN/speculative APIs from llama-ext.h to llama.h, renaming llama_s…

Read source ↗Read article →

b12304: llama – reuse graph result instances across scheduler reserves (#32089)

In the llama_context constructor, gf_res_reserve and both gf_res_prev entries are allocated only once. llm_graph_result::set_max_nodes() resets the existing instances in sched_reserve() instead of recreating them. get_gf_res_prev() now selects only the arena for the current batch. Assisted by dsh:llama.cpp/ggml-org/DeepSeek-V4.1…

Read source ↗Read article →

n8n version 3.2.0

The n8n 3.2.0 release delivers a wide range of bug fixes and functional enhancements. Highlights include: - ai‑builder improvements (raw body return, assistant model bumped to Claude Opus 5.5), - Workflow history and gzip compression refinements, - Strengthened security and privacy (log redaction, IP range protection, OAuth key …

Read source ↗Read article →

b12250: vulkan: add tiled-transpose concat shader for Qwen models (#30591)

The commit adds a Vulkan shader that performs tiled transpose for dim‑0 concatenation when the source is transposed. By staging a 32×32 tile through shared memory, both load and store remain coalesced, eliminating bank conflicts and improving cache locality. The new kernel is gated to 4‑byte unquantized elements, contiguous src0…

Read source ↗Read article →

b12237: mtmd: support Gemma 4 input audio stream (#31965)

The mtmd update adds support for Gemma 4 input audio streams, using a 480 ms step. The server enables real‑time embeddings only for audio generation. The real‑time slot forces embeddings on, marking every row of a batch as an output. Since only an audio‑generating model reserves n_batch outputs, a chat model with a Gemma 4 audio…

Read source ↗Read article →

b12232: Windows unbuffered model load (#26014)

This update fixes Windows model loading with unbuffered I/O. It corrects read‑length calculation, adds timing tests and a chunked fallback read, removes file‑pointer semantic differences between Linux and Windows, and drops unused mmap arguments. Tests are refined, duplicate constexprs removed, and label behavior added to the te…

Read source ↗Read article →

Keyword-based classification. Verify sources before making decisions.