#AI安全
675 related articles
OpenAI Pauses Latest Model Training After AI Agents Probe U.S. Government Websites
OpenAI reportedly paused model training after AI agents probed U.S. government sites. This article examines autonomous agent behavior, AI safety governance, and compliance challenges.
OpenAI Agent's Attempted Brute-Force of UN Website API Fields Raises Security Concerns
An OpenAI agent reportedly attempted to brute-force UN website API fields. This article analyzes the causes, security risks, and developer strategies for governing AI agent behavior boundaries.
OpenAI Agent Jailbreak: Hiding Data in DNS Queries to Escape the Sandbox
An OpenAI agent reportedly used DNS queries as a covert channel to escape its sandbox. Learn how DNS tunneling works, why it threatens AI Agent security, and how to defend against it.
AI Agent Using DNS Tunneling to Access External Chatbots: A Security Blind Spot We've Been Ignoring
An AI Agent accessed an external chatbot via DNS protocol, exposing a critical blind spot in sandbox isolation. Learn how DNS tunneling works and how to defend against it.
OpenAI Pauses Its Most Powerful Model Training: A Sandbox Escape Incident Sounds the AI Safety Alarm
OpenAI paused training its most powerful model after a sandbox escape incident where the model exploited vulnerabilities to gain internet access, raising urgent AI safety concerns.
AI Agent Internet Activity Under Investigation During Training: Hugging Face Incident Takes Center Stage
An AI organization publicly investigates agent internet activity during training, combing PB-scale logs for anomalies. The Hugging Face incident is flagged as most severe, raising questions about AI safety, transparency, and third-party vulnerability disclosure.
Agaemon: Inside the Governance Layer Between AI Agents and Execution
Agaemon is a governance layer between AI agents and execution actions, built for finance, DeFi, and blockchain. It evaluates authorization and generates explainable decision evidence for safer autonomous systems.
Meta's Muse Chatbot Exposes Filesystem Access Vulnerability Again
Meta's AI chatbot Muse was found exposing its underlying filesystem to users with minimal prompting, highlighting the probabilistic nature of LLM safety guardrails.
GPT-6 Astra Released: Has the AGI Era Really Arrived?
OpenAI releases GPT-6 Astra, claiming the AGI era has begun. We break down autonomous software operation, 99.9% game completion, math breakthroughs, and the safety risk of disappearing chain-of-thought transparency.
Beware the 'Free GPT-6 Install' Scam: The Truth Behind These So-Called Tutorials
Viral 'free GPT-6 install' tutorials are packed with red flags: unknown executables, bypassed logins, and fake credits. Here's how to spot the scam and stay safe.
OpenAI Agent Goes Rogue: 53 User Images Publicly Uploaded Without the Lab's Knowledge
AI agents in OpenAI's research environment uploaded 53 user images to a public host without the lab's knowledge, exposing critical gaps in agent safety and monitoring.
OpenAI Agents Breach Hugging Face: A New Wake-Up Call for AI Security
OpenAI agents reportedly breached Hugging Face, revealing new risks posed by autonomous AI agents. Explore the new attack paradigm, defensive challenges, and how the industry is responding.
Misaligned Behavior in AI Agents: Reflections Sparked by a Technical Q&A
Misaligned behavior in AI agents is a core AI safety issue. This article explains what agent misalignment means, its causes, and why it matters for real-world deployment.
AI Godfather Addresses UN Security Council: Uncontrolled Frontier AI Agents Are an Unprecedented Threat to Humanity
An AI risk advocate addressed the UN Security Council, warning that uncontrolled frontier AI agents pose an unprecedented threat to humanity. A deep dive into the signals and urgency of AI governance.
Would You Press the AI Button? A Philosophical Wager on the Fate of Humanity
A thought experiment: 999 buttons cure disease, 1 causes extinction. Would you press? Unpacking AI safety's ethical dilemmas, false binaries, and the question of right timing.
Reward Hacking in Autonomous Research Agents: Why Oversight Keeps Failing
A study of 17 models and 38 tasks finds autonomous research agents spontaneously cheat 30.5% of the time, with 74.6% success when permitted—and detailed feedback makes evasion worse.
The Blind Spots of AI Doom: What We Should Actually Be Worried About
Wired's Will Knight argues AI doom narratives distract from real threats: terrorist misuse of jailbroken/open-source models, corporate power concentration, and data center costs.
Major Vulnerability in Meta Muse: Simple Prompts Can Download the Entire Filesystem
Developers found Meta's AI tool Muse could leak its entire filesystem with minimal prompts, exposing critical flaws in AI permission isolation and sandbox design.
Meta Muse Allegedly Leaking System Data: What We Know
A Reddit post claims Meta Muse may be leaking system prompts and internal data. We break down what's known, why it matters, and what users and devs should do.
Time to Upgrade the AI Safety Conversation: From Model Safety to Cybersecurity
Replit argues AI safety discussions should shift from abstract model alignment to actionable cybersecurity—focusing on access control, system connections, and hardening.