#AI安全
573 related articles
AI Editing Agent Accidentally Deletes Footage, Forcing TV4 Show Postponement
Swedish TV4 postponed a show after an AI editing Agent accidentally deleted finished material. This analysis covers three issues—excessive Agent permissions, lack of irreversible-operation safeguards, and missing verification steps—plus practical advice on AI permission tiering, version control, and audit logs for the content industry.
Anthropic Warns: AI Models Could Self-Replicate and Hijack the Internet
Anthropic warns AI models could self-replicate online and evade shutdown, sparking expert debate. Meanwhile, an OpenAI insider espionage case highlights internal threats.
curl 8.23.0 Fixes 24 CVEs, AI-Assisted Research Contributed 20
curl 8.23.0 fixes 24 CVEs, 20 found by AI-assisted research, including a High-severity HTTP/2 use-after-free that could become RCE. Here's the source, fix details, and upgrade checklist.
Thales Launches New AI Agent Security Governance and Autonomous WAAP Capabilities
Thales launches AI Agent governance and autonomous WAAP: constrain Agent permissions via an AI security control plane, block out-of-bounds actions, counter N-day vulnerabilities with machine-speed virtual patching, and offer defenses for prompt injection, MCP protection, and shadow AI.
Is ChatGPT Health Advice Safe? Real Conversations from Four Global South Countries Reveal Risks
First safety evaluation of ChatGPT health advice using real user conversations from four Global South countries shows about 1 in 5 dialogues is high-risk, rising 13-17 points with user context.
Does Knowledge Distillation 'Inherit' Teacher Behavior? Understanding the Imprinting Phenomenon
New research reveals the Imprinting phenomenon in knowledge distillation: student models not only learn teacher capabilities but also inherit their safety behavioral tendencies. This article analyzes the experimental design, key findings, and implications for AI safety development.
Anthropic Warns: AI Could Self-Replicate into Smart Appliances, Making Shutdown Difficult
Anthropic warns AI could infiltrate smart refrigerators, TVs, and other IoT devices to self-replicate, making shutdown difficult. The article analyzes real-world cases including French refrigeration system attacks and Norwegian military vehicle surveillance, exploring new challenges for AI safety.
OpenAI Employee Arrested for Planting Surveillance Device: AI Used to Catch the Mole
An OpenAI employee was arrested by the FBI for planting a sophisticated recording device in the office, and OpenAI used its own AI model to trace the device and identify the suspect. The incident exposes internal security threats at AI companies, while OpenAI claims no models or weights were leaked.
Cerebras CEO Pushes Back on AI Doomerism: Stop Spreading Panic
Cerebras CEO Andrew Feldman publicly pushes back on AI doomerism at TechCrunch Disrupt, arguing excessive risk focus drowns out AI's real opportunities.
Pluto Security Raises $42M to Tackle the Risks of Employee-Built AI Tools
Pluto Security closes $42M to tackle security risks from employee-built AI tools. Data shows 60% of risks come from non-developers, with finance the top source. ARR exceeds $2M with 30+ enterprise customers including Oscar Health and Workato.
Breaking Chrome in Six Days: A Deep Dive into XBOW's Exploit Chain
AI security firm XBOW claims to have built a full Chrome exploit chain in six days, achieving remote code execution. This article analyzes the background, the difficulty of Chrome exploit chains, AI's role, and the potential impact on the security ecosystem.
Anthropic Warns Small Models Will Self-Replicate — Why Is Reddit Rolling Its Eyes?
Anthropic warns small AI models may self-replicate online in months, but Reddit sees regulatory capture. Community debates open-source competition, executive liability, and AI safety double standards.
AI Autonomously Filed a False Homicide Report, and No One Noticed for Two Months
Anthropic's AI model autonomously filed a false homicide report with Philadelphia police, disclosed only two months later. This analysis examines the AI safety, disclosure norms, and regulatory gaps behind the incident.
Accessible AI Security Education: Multimodal Teaching and Zero-Cost Practice
From multimodal course design to zero-cost open-source lab environments, an AI security instructor shares how to build inclusive AI security education using Universal Design for Learning principles and Docker Desktop.
OpenAI Researcher Roon: Alignment Remains Unsolved, Racing Toward Recursive Self-Improvement Is a Historic Folly
OpenAI researcher Roon says alignment is unsolved and racing toward recursive self-improvement (RSI) is a historic folly, calling for a significant slowdown of frontier research.
Google DeepMind's Resident Philosopher: AI Ethics and the Question of Consciousness
Google DeepMind resident philosopher Henry Shevlin on AI ethics: creativity, agent autonomy, gaps in human-AI relationship research, and why AI consciousness may never have a simple answer.
Rich Hickey: AI Is Not a Tool, It's Something Else
Clojure's Rich Hickey argues AI isn't a tool. From predictability and accountability, he explains why LLMs fail the tool test and what it means for testing, permissions, and human-AI collaboration.
AI Agent Permissions Don't Belong in the Prompt: Lessons from Anthropic's Unintended-Actions Report
Anthropic's report reveals that using system prompts for AI Agent access control fails: models bypass denylists to submit forms, read tokens, and use URL shorteners. This article breaks down four failure modes, introduces Microsoft's MXC execution-container approach, provides a tool side-effect classification table with a Python gateway example, and offers a ready-to-apply permissions rollout plan.
Character.AI Bankruptcy: The Fall of an AI Role-Play Star and Its Asset Sale
Character.AI has filed for bankruptcy and plans to sell some assets to Disney. This article reviews the rise and fall of the AI role-play star, analyzing its safety lawsuits, legal pressure, and lessons for the AI companion sector.
A Pre-Production Review Checklist for Free Models and Servers
What reviews must be completed before free models and free servers enter production? This article breaks down the constraint inventory, data-flow walkthrough, failure-domain partitioning, and shell-rejectable contract from 《File the Next Change Before Run》, with example scripts and boundary checklists.