#model poisoning
58 related articles
·3 min
Claude Code in Practice: How AI Coding Tools Can Protect Your Secret Keys
When AI coding assistants have full filesystem access, are your API keys and cloud credentials safe? Deep dive into secret leakage risks and practical defenses.
Read more →
How a Single Word in a Prompt Can Shap…
·2 min
How a Single Word in a Prompt Can Shape AI Output: The Well of Possibilities Explained
How does a single word in a prompt affect LLM output? This deep dive explains autoregressive generation, probability shifts, and practical tips for neutral prompting.
Read more →
·3 min
jscrambler npm Package Compromised: Supply Chain Attack Incident Response and Protection Guide
The jscrambler npm package v8.14.0 was hit by a supply chain attack, with malicious code able to steal environment variables, CI secrets, and deployment credentials. This guide covers self-inspection, credential rotation, and dependency hardening.
Read more →
·6 min
GPT-5.6 vs Grok 4.5 on the Same Day: A Packed Week of Flagship AI Model Launches
OpenAI's GPT-5.6 launches with Sawa, Terra, and Luna sub-models the same day as Musk's Grok 4.5, while Anthropic, Meta, and NVIDIA make their moves. A packed week of flagship AI launches.
Read more →
·6 min
AI Agent Security in Practice: 7 Attack Techniques and a Five-Layer Defense System
An in-depth breakdown of the 7 major attack techniques against AI agents (prompt injection, data poisoning, image attacks, etc.) and a five-layer defense system, with real cases from Doubao and DeepSeek.
Read more →
Anti-AI Fonts: Adversarial Typography …
·3 min
Anti-AI Fonts: Adversarial Typography That Humans Can Read but Machines Can't
Anti-AI fonts use adversarial glyph design to exploit the perceptual gap between human and machine vision, blocking OCR and LLM scraping. A deep dive into the tech, limits, and data sovereignty implications.
Read more →
·5 min
Crew: Enabling Multiple AI Agents to Collaborate and Share Knowledge Like a Team
Crew is an open-source AI agent collaboration framework whose core idea is to build a "Stack Overflow" for agents—letting multiple agents share experience and accumulate knowledge, shifting from optimizing single agents to building evolving teams.
Read more →
·4 min
Surge in Claude Mythos Preview Vulnerabilities: How Can We Ensure LLM Security?
After the release of Claude Mythos Preview, critical security vulnerabilities surged, raising widespread concern. This article analyzes the tension between rapid iteration and security, explores LLM attack surface challenges, and offers practical defense strategies.
Read more →
·4 min
AIML Capstone Project Selection: Is a Multi-Agent Debate System Worth Building?
Torn over your capstone topic? This article analyzes the academic value, feasibility, and innovation potential of a Multi-agent Debate system to help AIML students decide.
Read more →
·5 min
Prismata Analysis: How to Defend Web Agents Against Cross-Site Prompt Injection Attacks
Cross-site prompt injection is becoming the trickiest security threat for Web agents. This article analyzes the Prismata project's 'confining defense' approach—controlling injection's blast radius via context isolation, permission boundaries, and trust grading.
Read more →
·2 min
Open Source as Critical Infrastructure: How Should Governments Spend Their Money Effectively?
Governments frequently label open source software as critical infrastructure, but the real test is whether budgets flow to maintainer time, infrastructure, and incident response—not just surface-level audit reports.
Read more →
LLM Security Benchmarking: Current Sta…
·4 min
LLM Security Benchmarking: Current State, Challenges, and Practical Guide
Why is it so hard to establish unified LLM security benchmarks? This article analyzes core challenges in LLM security evaluation—covering jailbreaks, prompt injection, red teaming, and more—with practical strategies for developers.
Read more →
·7 min
Demystifying Claude Code's Harness in 60 Lines of Python: The Core of AI Coding Tool Performance Differences
The same model scores 77% in Claude Code but jumps to 93% in Cursor—the only variable is the Harness. This article dissects how AI coding tools work in 60 lines of Python.
Read more →
·6 min
Unsloth Major Update: DeepSeek-V4 Support and a Deep Dive into NVFP4 Quantized Export
Unsloth v0.1.481-beta adds full DeepSeek-V4-Flash support, NVFP4/FP8/imatrix GGUF quantized export, 1.3x faster GRPO, 3-5x faster MoE training, and an OpenAI-compatible API service in Studio.
Read more →
·5 min
6 Self-Hosted AI Agents Compared: Which One Is a Truly Self-Evolving Digital Employee?
An open-source AI Agent with 380K stars ranks only third? This comparison of 6 self-hosted AI Agents scores them on persistence, self-evolution, and data control—revealing why Generic Agent won with just 3,000 lines of code.
Read more →
·7 min
In-Depth Analysis of the GitLost Vulnerability: How GitHub AI Agents Were Tricked into Leaking Private Repositories
Security researchers disclose the GitLost attack, which uses prompt injection to trick GitHub AI agents into leaking private repository source code. An in-depth analysis of the attack principle, AI agent security risks, and developer defense recommendations.
Read more →
·6 min
Alibaba Bans Claude Code: The Enterprise Security Red Line for AI Coding Tools
In-depth analysis of Alibaba's comprehensive internal ban on Claude Code: from the hidden-marker controversy and Anthropic's regional-restriction stance to five core questions of enterprise AI coding tool security admission.
Read more →
·4 min
Alibaba Plans to Ban Claude Code: The Battle Between Backdoor Risks and Enterprise AI Security
Alibaba reportedly plans to ban Claude Code internally over backdoor and data leakage concerns. A deep dive into enterprise AI security, supply chain trust issues, and what it takes for AI tools to win enterprise adoption.
Read more →
·3 min
Unsloth v0.1.46-beta Update: DiffusionGemma Tool Calling Disabled and Artifacts Canvas Enabled by Default
Unsloth v0.1.46-beta is out with key DiffusionGemma changes: tool calling disabled by default, artifacts canvas enabled. A deep dive for LLM fine-tuning devs.
Read more →
·6 min
Claude Code Chinese Guide: A Complete Path from Zero to an Efficient Workflow
A systematic Claude Code learning guide built for Chinese developers, covering ten core modules including Slash Commands, Memory, MCP, and Hooks, with a three-tier path to build an AI coding workflow in 11–13 hours.
Read more →