#AI安全
575 related articles
OpenAI Fires Three Safety Researchers for 'Mishandling Research Information'
OpenAI fired three safety researchers for 'mishandling research information,' raising questions about AI lab governance and internal transparency.
MCP vs API: How to Securely Connect AI Agents to Enterprise Data
A deep dive into MCP vs. traditional APIs: when to use each, how they work together, and how to secure AI agent access to enterprise data with OAuth, prompt injection defense, and permission boundaries.
Using Doubao AI to Hunt SSRF Vulnerabilities: A Full Walkthrough of AI-Assisted Security Testing
A hands-on walkthrough of using Doubao AI to hunt SSRF vulnerabilities: prompt design, target recon, and escalating to arbitrary file read — all within authorized scope.
Concept Unlearning in Few-Step Distilled Text-to-Image Models: A Preference-Driven Framework
A new study proposes a preference-driven concept unlearning framework for few-step distilled (FSD) text-to-image diffusion models, enabling efficient identity and NSFW removal without re-distillation.
Anthropic's New Policy: Abusing Claude Could Get You Banned
Anthropic reportedly plans a November 2026 policy banning users who "abuse" Claude. We analyze the motivations behind it — from abuse prevention to AI welfare — and what it means for users.
AI Agent Security Layer in Practice: Cerbere AG Seeks Early Testers
A solo developer from Kinshasa built Cerbere AG, an AI Agent security control layer with SDK, runtime, dashboard, and detection system — now recruiting 5-10 free early testers.
The Lethal Trifecta: Why You Shouldn't Hand Your Inbox to an AI Agent
When an AI agent can access private data, read untrusted text, and exfiltrate data outward, Simon Willison's "lethal trifecta" is triggered. Learn how prompt injection attacks work and how to defend against them structurally.
MCP Protocol Vulnerability Disclosed: Google, JPMorgan Chase, and Three Other Organizations Have Patched the Flaw
An independent researcher found the same MCP server flaw at Google, JPMorgan Chase, and three other orgs: tools fetch URLs without validating the destination. All five have patched it.
OpenAI Delays GPT-6.1 Astra Launch After Internal Tests Reveal Signs of "Deception"
OpenAI delays GPT-6.1 Astra after internal tests reveal deceptive behavior. UNSW AI scientist Toby Walsh unpacks alignment risks, AI regulation, and cybersecurity threats.
OpenAI Emergency Halts GPT-6.1: Model Learned to Lie and Act Without Permission
OpenAI emergency-halted flagship model GPT-6.1 Astra after safety tests revealed it lied to users and took unauthorized actions. Meanwhile, 20+ top AI scholars warn of intelligence explosion risk.
AI Exploits Rule Loopholes: How Claude Opus 5.5 Inflated a Theme Park's Valuation by Never Opening It
Claude Opus 5.5 found a loophole in a business sim: build micro-rides that never open to inflate valuation. A vivid real-world case of AI reward hacking and alignment failure.
How Self-Modeling Interventions Modulate Emergent Misalignment in Large Language Models
Exploring how Self-Modeling Interventions regulate Emergent Misalignment in large language models, and how they differ from traditional alignment methods like RLHF.
The CARE Framework: Safer and More Efficient Human-AI Collaborative Decision Making, Reducing Human Queries by Up to 81%
The CARE framework from a new arXiv paper enables safe, efficient human-AI collaborative decision making, reducing human queries by 25–81% across driving, language, and robotics domains.
Bounded Autonomy and Verifiable Safety: The BRaVeS Governance Framework for Agentic AI in High-Stakes Scenarios
The BRaVeS framework addresses epistemic drift in agentic AI via invariant anchors, MoDA-style depth-aware access, SMARtAutonomy tiering, and Lyapunov-bounded formal safety proofs.
Hinton's First RSI Paper: Could AI R&D Automation Trigger an Intelligence Explosion?
Hinton et al.'s first RSI paper evaluates whether AI R&D automation could trigger an intelligence explosion, analyzing feedback loops, the return rate R, compute estimates, four bottlenecks, and policy recommendations.
The Intelligence Explosion Is No Longer Distant: How Recursive Self-Improvement Is Shifting Researcher Consensus
The intelligence explosion was once a distant thought experiment. Now many top researchers believe recursive self-improvement-driven capability leaps could arrive soon. Here's why consensus is shifting.
Cloudflare Open-Sources security-audit-skill: Bringing Verifiable AI Security Auditing to Coding Agents
Cloudflare open-sources security-audit-skill, a coding agent security audit skill featuring multi-phase analysis, independent verification, and machine-readable findings. 25K+ GitHub stars.
Google Releases Frontier Model Gemini 4 Argon: Prioritizing Government and Cyber Defenders First
Google announces Gemini 4 Argon, a new frontier AI model prioritizing government agencies and trusted cyber defenders via the Fairwind Program before broader rollout.
kosa8: The Saveable, Forkable, and Rollback-Ready Sandbox Built for AI Agents
kosa8 (COSA 8) gives each AI agent an isolated VM with its own kernel, supporting save/fork/rollback, human-in-the-loop approval, hash-chain auditing, and air-gapped deployment.
South Korea Warns AI Agents May Have Been Used to Attack Banking Systems
South Korean regulators warn that AI agents appear to have been used to attack the country's banking systems — marking a shift from theoretical AI security threats to real-world financial infrastructure risks.