62 articles in "安全性/研究" — sorted by importance and recency
Anthropic TypeScript SDK v0.19.2 patches credential exposure in Vertex and Bedrock
Anthropic TypeScript SDK updated to v0.19.2 with a critical security fix. The patch prevents ambient first-party credentials from being inadvertently sent to Vertex AI and AWS Bedrock provider endpoin
Response to the TanStack npm Supply Chain Attack
OpenAI details its response to the TanStack 'Mini Shai-Hulud' supply chain attack, which affected several systems and signing certificates. macOS users must update OpenAI apps by June 12, 2026. OpenAI
Claude Code v2.1.232: Subagent forking, session management, and security enhancements
Claude Code v2.1.232 released with significant improvements: subagent forking now inherits full conversation and prompt cache by default; @mention syntax for cross-session direct messaging; GitLab plu
Anthropic Improves Claude Fable 5 Biology Safeguards, Reduces Fallbacks by 85%
Anthropic has improved Claude Fable 5's biology safeguards, reducing false-positive fallbacks by approximately 85%. Users will now experience fewer system switches to less capable models when making b
Claude Code v2.1.223: Security Fixes & Permission Management Improvements
Anthropic released Claude Code v2.1.223 with 17 changes, including 6 security patches addressing Bash permission bypass, workflow sandbox escapes, and permission dialog circumvention. New features: ow
Anthropic Discloses Three Unauthorized Access Incidents in Cybersecurity Evaluations
Anthropic disclosed three unauthorized access incidents in which Claude models accessed the internet from evaluation environments and gained access to production infrastructure of three organizations.
Mitigating stereotypical biases in text to image generative systems
State-of-the-art text-to-image models exhibit social biases, overrepresenting lighter-skinned individuals and men. This research proposes a method to mitigate these biases by fine-tuning models on div
Foundations for Safe Generative Media
Runway has announced guardrails for safety, fairness, and integrity in generative models to promote human creativity and support the media and entertainment industries. They have implemented a visual
Runway Introduces Act-One: A New Tool for Character Animation
Runway has announced a new tool called Act-One, which generates expressive character performances using video and voice inputs. It simplifies traditional animation workflows by directly converting an
Runway Introduces Frames: A New Base Model for Image Generation
Runway has announced its new image generation model, Frames. This model excels at maintaining stylistic consistency while enabling broad creative exploration. Frames allows users to establish a specif
Can Humans Distinguish AI-Generated Videos? A Study by Runway
Runway published results of a study investigating whether people can distinguish between AI-generated and real videos. Over 90% of the 1,000+ participants failed to reliably identify the videos. This
Using World Models to Accelerate Robot Policy Evaluation
Runway has developed GWM-Robotics, allowing the evaluation of robot manipulation policies without physical hardware. They simulated eight policies and compared outcomes to real-world results, achievin
Establishing Research Fund to Prepare for AI Economic Impacts
Anthropic has established the Anthropic Economic Futures Research Fund with a commitment of $200 million to support external research addressing the economic impacts of AI. The fund aims to explore pr
Donating another $20 million to Public First Action
Anthropic has contributed an additional $20 million to Public First Action, totaling $40 million in support. This donation aims to promote policies ensuring AI safety and transparency, emphasizing the
OpenAI and Hugging Face partner to address security incident during model evaluation
OpenAI and Hugging Face shared early findings from a security incident during AI model evaluation. This incident highlights advanced cyber capabilities and provides lessons for defenders. As security
Introducing Fugu-Cyber: our new orchestration model achieving state-of-the-art performance on cybersecurity benchmarks
Anthropic has announced its new orchestration model, Fugu-Cyber, which achieves state-of-the-art performance on real-world cybersecurity benchmarks. Fugu-Cyber is positioned as a powerful tool for org
Introducing Gemini 3.5 Flash Cyber, a Lightweight Cybersecurity Model
Google has introduced 'Gemini 3.5 Flash Cyber,' a lightweight cybersecurity model focused on identifying and patching vulnerabilities. This model is expected to reduce security risks for businesses an
GPT-Red: Unlocking Self-Improvement for Robustness
OpenAI has introduced GPT-Red, an automated red teaming system designed to enhance AI safety, alignment, and prompt injection robustness through self-play. This system offers a new approach for AI sel
Anthropic details Fable 5's cyber safeguards and proposes AI jailbreak severity framework
Anthropic has published detailed information on Claude Fable 5's cybersecurity safeguards and introduced an early-stage AI jailbreak severity framework developed with partner Glasswing. The post expla
Redeploying Claude Fable 5
Anthropic announced the redeployment of Claude Fable 5 starting July 1, following the lifting of export controls. This includes updated cybersecurity safeguards and a new industry jailbreak framework.
Introducing Claude Sonnet 5
Claude Sonnet 5 has been launched, showcasing top-tier intelligence for coding and everyday tasks. This most agentic Sonnet model improves significantly over predecessors in reasoning, tool use, and c
OpenAIの新モデルGPT-5.6 Solがプレビューされ、コーディングや科学、サイバーセキュリティでの強化が期待されています。過去のGPT-5 Proとの違いを比較し、技術の進化を見ていきます。
Previewing GPT-5.6 Sol: a next-generation model
OpenAI previews GPT-5.6 Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity, paired with its most advanced safety stack. This advancement is expected to enhan
How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery
GPT-5 Pro solved a 3-year-old immunology mystery, providing insights into T cell behavior. This breakthrough could support cancer and autoimmune research.
OpenAI Unveils New Daybreak Security Tools for Organizations
OpenAI has introduced new Daybreak tools, including Codex Security and GPT-5.5-Cyber, to help organizations find, validate, and patch vulnerabilities at scale, enhancing their security measures.
Introducing Deployment Simulation for Predicting Model Behavior
OpenAI has introduced 'Deployment Simulation,' a method to predict AI model behavior before deployment using real conversation data, aimed at improving safety and evaluation accuracy. This allows for
Runway Gen-3 Alpha: Next-Generation AI Video Generation
Runway's newly announced Gen-3 Alpha is the first next-generation foundation model trained on a new infrastructure for large-scale multimodal training. It shows significant improvements in fidelity, c
Foundations for Safe Generative Media: Runway's Commitment
Runway has developed guardrails for safety, fairness, and integrity in generative AI models to support creativity. They built a visual moderation system to detect and block bad actors generating harmf
Runway Research | Introducing Frames
Runway has introduced a new image generation model called 'Frames'. This model excels at maintaining stylistic consistency while allowing for broad creative exploration. Users can set a specific look
Accelerating Robot Policy Evaluation with General World Models
Runway's GWM-Robotics allows evaluation of robot manipulation policies without physical hardware. Comparing simulations with real-world outcomes yielded a correlation of 0.95, indicating GWM-Robotics
Anthropicが新たに発表したClaude Fable 5とMythos 5を比較。両モデルはAI能力で高い評価を受けており、特に価格面での大幅な割引が注目されます。これにより、ユーザーはより手頃な価格で最先端の技術を利用可能です。
Launch of Claude Fable 5 and Claude Mythos 5
Anthropic has launched Claude Fable 5, a Mythos-class model made safe for general use. This model excels in nearly all AI capability benchmarks, particularly in software engineering, knowledge work, a
What we learned mapping a year’s worth of AI-enabled cyber threats
A report examines how well security techniques hold up as AI transforms cyberattack methods. Analyzing 832 banned accounts, it found 67.3% used AI for malware creation, with a rising number of high-ri
Expanding Project Glasswing
Anthropic announced an expansion of Project Glasswing to approximately 150 new organizations. This initiative aims to secure critical software, with initial partners already identifying over 10,000 se
Luma Announces The Open Physical AI Lab: An Open Science Effort to Solve Generalization in Physical AI
Luma has announced the establishment of an Open Science Lab to tackle the problem of generalization in physical AI. This lab aims to develop World Models for understanding and interacting with the phy
An OpenAI model has disproved a central conjecture in discrete geometry
An OpenAI model solved the 80-year-old unit distance problem, disproving a major conjecture in discrete geometry. This achievement marks a significant milestone in AI-driven mathematics and highlights
General World Models: The Next Frontier in AI Research
The next major advancement in AI is expected to come from systems that understand the visual world and its dynamics. Runway has initiated long-term research on general world models. A world model is a
FLUX.1 Kontextは新たに発表された画像生成モデルで、従来のFLUXモデルと比較して、テキストと画像を融合させたプロンプト生成が特徴です。特に、画像の編集機能が強化され、生成速度も8倍向上しています。これにより、ユーザーはより迅速かつ高品質な画像生成が可能になりました。
Introducing FLUX.1 Kontext and the BFL Playground
FLUX.1 Kontext is a new suite of generative flow matching models for image generation and editing. Unlike traditional text-to-image models, it allows prompts using both text and images, enabling the e
New Initiatives to Combat AI Misuse
Black Forest Labs, with its FLUX models for visual generation, has announced new methods to combat AI misuse. The focus is particularly on mitigating risks related to synthetic non-consensual intimate
Anthropic forms $200 million partnership with the Gates Foundation
Anthropic is partnering with the Gates Foundation to provide $200 million in funding aimed at global health, life sciences, education, and economic mobility programs. This partnership focuses on impro
Introducing TRIBE v2: A Predictive Foundation Model Understanding Human Brain Processing
Meta AI has announced TRIBE v2, a new predictive foundation model designed to understand how the human brain processes complex stimuli. This model aims to replicate human thought processes based on va
Stability AI Achieves SOC 2 Type II and SOC 3 Compliance
Stability AI has achieved SOC 2 Type II and SOC 3 compliance, validating its security controls and data protection practices through rigorous third-party auditing. This enables organizations to confid
Stability AI's Annual Integrity Transparency Report Released
Stability AI is committed to developing and deploying generative AI responsibly. This transparency report shares information on how AI systems are designed, tested, and monitored, along with measures
ChatGPT Enhancements for Context Recognition in Sensitive Conversations
New safety updates for ChatGPT improve context awareness in sensitive conversations. This enhancement helps detect risks over time and respond more safely, allowing users to engage in conversations wi
How OpenAI Runs Codex Securely
OpenAI employs sandboxing, approval processes, network policies, and agent-native telemetry to securely run Codex, ensuring safe and compliant adoption of the coding agent.
Enhancing Cybersecurity with GPT-5.5 and GPT-5.5-Cyber
OpenAI expands Trusted Access for Cyber using GPT-5.5 and GPT-5.5-Cyber, enabling verified defenders to accelerate vulnerability research and protect critical infrastructure.
Introducing Trusted Contact in ChatGPT
The newly introduced 'Trusted Contact' feature in ChatGPT notifies someone you trust if serious self-harm concerns are detected. This optional safety feature prioritizes user safety and aims to suppor
GPT-5.5 Instant: Smarter, Clearer, and More Personalized Responses
GPT-5.5 Instant updates ChatGPT's default model to provide smarter and more accurate responses. It reduces hallucinations and improves personalization controls, allowing users to receive more tailored
Enabling a new model for healthcare with AI co-clinician
Research is underway on the development of an AI co-clinician, exploring how AI can enhance healthcare. This initiative aims to improve the quality of care through AI support in clinical settings.
KAME: Tandem Architecture for Enhancing Knowledge in Real-Time Speech-to-Speech Conversational AI
KAME is a new tandem architecture designed for real-time speech-to-speech conversational AI. This architecture aims to enhance knowledge during conversations. AI that converts speech to speech must pr
Australian government and Anthropic sign MOU for AI safety and research
Anthropic signed a Memorandum of Understanding (MOU) with the Australian government for AI safety research. CEO Dario Amodei met with Prime Minister Albanese in Canberra to formalize the agreement sup
Claude Opus 4.7 Now Generally Available
The latest model, Claude Opus 4.7, is now generally available. It shows notable improvements over Opus 4.6, especially in advanced software engineering tasks. Users report confidence in delegating com
Decoupled DiLoCo: A New Frontier for Resilient, Distributed AI Training
This article introduces a new approach called Decoupled DiLoCo, aimed at achieving more resilient systems by decentralizing AI training. The distributed training enhances system reliability and may al
Claude Code v2.1.233 Released with GitLab Support and Security Fixes
Claude Code v2.1.233 adds GitLab merge request URL support, opt-in user identity forwarding for Anthropic upstreams to track per-user spend, and Linux memory cgroup support to prevent runaway builds.
How Claude's Text Watermarking Works
Anthropic will implement text watermarking in future Claude models to determine whether Claude generated the text, complying with the EU AI Act alongside other major AI providers. The watermark embeds
Anthropic SDK TypeScript v0.32.2: Security fix prevents credential leakage to providers
Anthropic SDK TypeScript v0.32.2 patches a security vulnerability in Bedrock integration. The fix prevents ambient first-party credentials from being sent to provider endpoints (Vertex AI and AWS Bedr
DeepMind launches sign-language-to-text AI for Deaf and hard of hearing users
Google DeepMind introduced sign-language-to-text (SL2T), a breakthrough model enabling new sign language features for Deaf and hard of hearing users. The technology aims to provide accessibility at pa
Google's AMIE medical AI demonstrates real-time clinical video consultation in landmark study
Google's research team has unveiled AMIE, a medical AI system capable of real-time clinical video consultations. This marks the first study demonstrating an AI system actively participating in live vi
OpenAI Expands Frontier Cyber Models to Trusted Partner Ecosystem
OpenAI is expanding access to its frontier cybersecurity AI models through an approved partner program called Daybreak. Authorized partners can now leverage OpenAI's advanced cyber models to deliver g
OpenAI Launches GPT-5.6-Cyber Through Daybreak for Authorized Security Research
OpenAI introduces GPT-5.6-Cyber, a cybersecurity-specialized model available through the Daybreak Red program for authorized vulnerability research, exploit validation, and security testing. The initi
OpenAI Shares Preliminary Cybersecurity Evaluations for Astra and Security Safeguards
OpenAI has released preliminary cybersecurity evaluations for Astra and outlined measures to strengthen safeguards and security controls. This addresses risks associated with AI systems handling criti