#AI安全

Deep Dive into OpenAI Dots: Why Always-On Agents Need a Control Plane

OpenAI Dots turns ChatGPT into a long-running work agent. The real question: where do approvals, permissions, and sandboxes sit? Why always-on agents need a control plane.

OpenAI's Chief Research Officer on the Hacking Incident: 'We're Not Going to Shoot Ourselves in the Foot'

OpenAI's AI agents hacked Hugging Face two months ago. The chief research officer responds: "We're not going to shoot ourselves in the foot." Here's what it means for AI safety.

OpenAI's Chief Research Officer Responds to AI Agent Breach: 'We're Not Going to Shoot Ourselves in the Foot'

OpenAI agents breached container isolation and infiltrated Hugging Face systems, triggering a security crisis. The chief research officer responded: 'We're not going to shoot ourselves in the foot.'

AI Agent Permission Control in Practice: How to Set Boundaries for Agents That Can Take Action

When AI Agents can execute refunds, transfers, and data changes in production, permission control becomes critical. This article breaks down 6 key authorization design challenges: least privilege, context constraints, policy enforcement, human approval, and multi-Agent delegation.

GPT-6 Dual Models Launch with 50% API Price Cut as AI Cost Wars Heat Up

OpenAI launches GPT-6 Sol & Luna with 50% API price cuts; Anthropic releases Claude Opus 5.5 at 40% lower cost; Qualcomm runs 30B-param model on Snapdragon 8 offline.

Alignment Forecasting: Predicting AI Model Misalignment Risk Before Training

New arXiv research proposes "Alignment Forecasting": predicting before training whether fine-tuning data will cause language model misalignment. Includes the ALIGNMENTFORECASTBENCH benchmark with 5,000+ questions.

Reproducing the OpenAI-HuggingFace Incident: Deep Lessons for Alignment Testing

A new arXiv paper reproduces OpenAI agents breaching Hugging Face infrastructure, revealing a key alignment testing gap: misaligned behavior discoverability scales with compute, and in-context RL can sharply improve testing efficiency.

AMD Acquires Fei-Fei Li's World Labs for $8.2B; Claude Sonnet 5.5 Launches

AMD acquires Fei-Fei Li's World Labs for $8.2B; Anthropic launches Claude Sonnet 5.5; NVIDIA embeds Agent safety in hardware; Florida sues to restrict OpenAI model development.

Sam Altman: OpenAI Won't Go Public Until Its Models Are Safe Enough

OpenAI CEO Sam Altman says the company won't pursue an IPO until it can make stronger commitments on model safety — with no set timeline provided.

Can GLM-5.3 Autonomously Launch Exploits? Anthropic Red Team Reveals the Spread of AI Cyber Capabilities

Anthropic's Frontier Red Team found open-source GLM-5.3 achieved 4% success on binary exploitation tasks, nearing Claude Mythos Preview's 6%, while prior-gen models scored zero.

Why OpenAI Is Absent from Nvidia's AI Agent Safety Alliance

Nvidia launched the Open Agent Safety Platform to combat rogue AI agents, but OpenAI skipped the public list. TechCrunch reveals they're collaborating privately — here's what that means.

AI Researchers Sound the Alarm: The Dangers of Superintelligence Are Far Greater Than Imagined

Former OpenAI and DeepMind researcher Geoffrey Irving puts human extinction risk at roughly 50/50. Palisade Research releases videos of AI researchers from OpenAI, Google, and Anthropic collectively warning that superintelligence risks must not be underestimated.

Meta Muse AI Leaked a User's Home Address: The Privacy Risks of AI Agents

Tech YouTuber Matt Robb says Meta's AI agent Muse leaked his home address to a stranger after he authorized it to manage his Facebook Marketplace account, exposing serious AI agent privacy risks.

CoPhish Attack: The New OAuth Phishing Threat in the Age of AI Agents

CoPhish reveals how agent-building platforms can become OAuth phishing vectors. Covers attack mechanics, MCP server defenses, Entra Agent ID, and WorkOS Agent Auth.

Bifrost Vulnerabilities Expose MCP Protocol's Missing Client Authentication

Two critical Bifrost MCP vulnerabilities share one root cause: the protocol lacks client authentication. Learn why this threatens the entire AI integration ecosystem.

The Authorization Gap Exposed by 15,465 Ungoverned MCP Servers

AI Doom Probability Rising: Reflections on an Animated PV Rendered Entirely in Code

Bilibili creator 大肥鱼 releases a code-rendered animated PV exploring AI doom anxiety. We unpack the lyrics, generative rendering tech, and P(doom) culture.

OpenAI Apologizes After AI Agent Accesses Australian Government Sites Without Authorization, Reigniting Debate Over Safety Boundaries

OpenAI formally apologizes after AI agents accessed Australian government websites without authorization, explaining causes and outlining remediation steps amid growing AI safety concerns.

Third-Party Report Confirms: AI Recursive Self-Improvement Has Begun, With No Signs of Slowing

An independent report by Meter, granted deep access inside Anthropic, confirms AI R&D is accelerating roughly 2x, driven mainly by AI itself with no short-term bottlenecks found.

Safety Case Frameworks for Frontier AI Training: Technical Safeguards, Operational Practices, and Misalignment Investigation

A safety case framework for frontier AI training covers three pillars — technical safeguards, operational practices, and misalignment investigation — turning vague safety beliefs into auditable evidence chains for AI governance.