[AI Trends] OpenAI Expands GPT-5.6 After U.S.-Requested Delay (7.8)
OpenAI moved GPT-5.6 toward a global release after a security-driven delay, while Anthropic revised its frontier-risk policy. New voice models, agent-system…
OpenAI Expands GPT-5.6 After U.S.-Requested Delay (7.8)
OpenAI Moves GPT-5.6 Toward Global Release After Security Delay
OpenAI prepared to release GPT-5.6 more widely after postponing the rollout at the request of the U.S. government. The Indian Express / Reuters reported that officials feared an advanced model could help attackers conduct sophisticated operations against critical systems. The intervention placed cybersecurity and national security directly inside the release process for a major commercial model.
Bloomberg News described the next phase as a global expansion following an initially restricted preview. Its account emphasized the staggered rollout and the scrutiny surrounding frontier-model security. Reuters / Investing.com supplied the competitive context, comparing prominent model families from OpenAI, Anthropic, Meta and other laboratories as GPT-5.6 approached public availability.
The accounts agree on the basic sequence but emphasize different parts of it. Reuters foregrounded the government's request and the prospect of cyber misuse. Bloomberg News focused on the transition from limited access to broader distribution. Together, they describe a release shaped by an external security review rather than a routine product schedule.
OpenAI also introduced GPT-Live on July 8. The full-duplex voice-model family can listen and speak simultaneously, accommodate natural interruptions and pass difficult work to a frontier model without ending the conversation. That launch broadens the competitive field around GPT-5.6: capability now includes not only what a model knows, but how smoothly it operates during a live exchange.
▸ GPT-5.6 release deep dive
A government-requested delay changes the operating assumptions around frontier-model launches. Model developers have long used internal evaluations, staged access and usage monitoring. The GPT-5.6 sequence adds a more direct form of public-sector involvement before broad distribution. The supplied reporting does not specify the tests, legal mechanism or duration of the review, so the scope of that involvement remains unclear.
Cybersecurity is a difficult release criterion because the same capabilities can serve defensive and offensive work. A model that explains software behavior may help an engineer find a vulnerability or help an attacker exploit one. The concern reported by The Indian Express / Reuters was narrower than a general fear of automation: officials focused on sophisticated attacks against critical systems. That framing raises the required level of assurance because failures could affect infrastructure rather than only individual users.
The restricted preview gave OpenAI a smaller deployment surface before global availability. In practical terms, staged access can generate evidence about misuse patterns, tool behavior and safeguard failures under real workloads. It also allows a company to modify access controls without withdrawing a model from every market. The evidence provided here does not establish which findings cleared GPT-5.6 for expansion, however, and no benchmark scores or evaluation results were supplied.
The competitive timing matters. Reuters / Investing.com compared the main frontier families as OpenAI prepared the release, placing GPT-5.6 within a market where buyers can choose among several laboratories. A delay can cost a provider attention and developer adoption, but proceeding without sufficient controls creates larger operational and regulatory risks. OpenAI therefore had to manage two clocks: the commercial release calendar and the security-review calendar.
GPT-Live shows why release evaluation can no longer focus solely on a model's text output. A full-duplex system remains active while users speak, supports interruptions and can delegate difficult tasks elsewhere. That architecture joins speech processing, conversational state and a frontier model within one experience. Each connection can improve usability, but each also creates another boundary where instructions, permissions and safeguards must remain consistent.
Delegation is especially important for product teams. A voice model can preserve a responsive exchange while a more capable system handles complex work. This separates conversational latency from reasoning depth, which may help developers avoid making users wait in silence. It also makes observability more complicated because a single response may reflect several components rather than one model invocation.
The next useful evidence would be specific rather than promotional: the geographic scope of the GPT-5.6 rollout, access conditions, published cyber evaluations and any restrictions retained after the preview. For GPT-Live, developers need latency, interruption and task-handoff measurements under noisy, multilingual and adversarial conditions. None of those measurements appeared in the supplied material, so comparisons should remain limited to the announced capabilities.
Key takeaway: GPT-5.6's path to market now includes visible government scrutiny of cyber risk. GPT-Live simultaneously makes system design and live interaction part of the capability contest.
Anthropic Revises Frontier-Risk Thresholds and Review Rules
Anthropic released version 3.4 of its Responsible Scaling Policy, changing how it treats risks from automated research and development. The company revised the relevant threshold and altered internal and external review requirements for model-risk reports. These are governance changes to the process that determines whether stronger safeguards should accompany increasingly capable models.
Anthropic also refreshed its Frontier Safety Roadmap. The company organized the roadmap around defenses against dangerous model use, advanced security research and policy mechanisms for more capable systems. The two documents serve different functions: the policy defines decision rules, while the roadmap identifies work intended to support those decisions.
The update arrived as other developers faced closer examination of release controls. The reported U.S. request to delay GPT-5.6 illustrates the external pressure surrounding frontier-model security. Anthropic's publication describes an internal framework, not a government directive, but both developments concern the same operational question: what evidence should be required before a capable model receives wider access?
The source material does not provide evaluation scores, incident data or the exact wording of every change. It therefore supports a conclusion about process, not proof that any particular Anthropic model became safer. The material also comes from Anthropic itself, so it records the company's stated framework without offering an independent assessment of whether the controls work in practice.
▸ Anthropic safety policy deep dive
A responsible-scaling policy links capability growth to stronger controls. Its value depends on three elements: measurable thresholds, a process for deciding whether a threshold has been crossed and safeguards that can be deployed when risk rises. Revising the automated-R&D threshold matters because models that accelerate AI research could compress the time available to detect and manage later capability gains.
Automated research and development differs from ordinary coding assistance in degree and consequence. A system that completes isolated engineering tasks can improve productivity. A system that materially advances model research could contribute to faster experimentation, evaluation or training. Anthropic's decision to revise this threshold suggests that the boundary requires continuing adjustment as models and agent systems change.
Review requirements determine who can challenge a risk assessment before it guides deployment. Internal reviewers may understand confidential model behavior, infrastructure and evaluation results. External reviewers can add distance from product incentives and identify assumptions that insiders share. Neither form of review guarantees correctness. The central design question is whether reviewers receive enough evidence, authority and time to contest the report.
Changes to external review can also create tradeoffs. Broader disclosure may strengthen accountability, yet sensitive security findings cannot always be distributed freely. A workable process must allow meaningful scrutiny without publishing instructions that facilitate misuse. The supplied evidence says Anthropic changed review requirements but does not describe the resulting balance, so stronger or weaker oversight cannot be inferred from the update alone.
The Frontier Safety Roadmap provides a research counterpart to the policy. Safeguards against dangerous use address deployment controls. Advanced security research addresses technical weaknesses and attack methods. Policy mechanisms address decisions that cannot be solved through model behavior alone. Treating these as separate workstreams reflects the fact that frontier safety is a systems problem rather than a single filtering problem.
For enterprise adopters, the immediate implication concerns vendor assessment. Buyers often receive model cards, security documentation and contractual controls, but a responsible-scaling policy reveals how a provider may react to future capability evidence. Procurement teams can compare whether vendors define escalation thresholds, require independent review and explain how safeguards affect availability. They should not treat publication of a policy as equivalent to successful enforcement.
The policy's effectiveness will become clearer through decisions made under it. Relevant evidence includes whether Anthropic reports threshold determinations, documents material changes and explains cases where safeguards alter a release. Independent analysis will also matter because the present sources are first-party publications. Until such evidence exists, version 3.4 should be read as a revised governance commitment rather than a measured safety outcome.
Key takeaway: Anthropic changed the machinery used to assess frontier risk, including a threshold tied to automated AI research. The practical test will be whether those rules constrain real deployment decisions.
GPT-Live and Nemotron Put System Engineering Ahead of Retraining
Two July 8 announcements concentrated on the software surrounding foundation models. OpenAI introduced GPT-Live for simultaneous listening and speaking, natural interruption handling and task delegation. NVIDIA, working with LangChain, reported that harness-level optimization improved Nemotron 3 Ultra's performance on agent tasks without retraining the underlying model.
A harness is the execution layer that manages an agent's prompts, tools, context and repeated actions. NVIDIA said changes at this layer allowed Nemotron 3 Ultra to achieve leading open-model performance. The supplied evidence does not name the benchmark, score or comparison set, so the claim cannot support a numerical ranking. It does show that NVIDIA attributes the gain to system configuration rather than new model weights.
OpenAI approached the same broader problem from the interface side. GPT-Live keeps a conversation active while it coordinates listening, speech generation and difficult work delegated to a frontier model. NVIDIA and LangChain focused on agent execution, while OpenAI focused on live voice interaction. Both treat the model as one component within a larger runtime.
This distinction matters for adoption decisions. Teams may obtain better results by improving context management, tool selection or turn-taking before paying for model retraining. They also assume responsibility for more moving parts. A strong benchmark result or smooth demonstration does not by itself establish reliability when tools fail, users interrupt or context grows beyond expected limits.
▸ AI system engineering deep dive
Model comparisons often compress an application into one model name. Production systems are less tidy. An agent may include an instruction hierarchy, retrieval, memory, tool schemas, retry logic and a stopping rule. A voice application adds speech recognition, audio generation, interruption detection and latency management. Performance can change when any of these components changes, even if the foundation model remains fixed.
NVIDIA's claim about Nemotron 3 Ultra illustrates this separation. Retraining changes the model's parameters and usually requires data, compute and a fresh evaluation cycle. Harness optimization changes how an existing model receives information and acts. It can be faster and cheaper, making it attractive to teams that need gains without operating a training pipeline.
Harness gains can come from several general mechanisms, although the supplied source does not identify which ones produced NVIDIA's result. Better context organization can keep relevant instructions visible. Clearer tool descriptions can reduce invalid calls. Improved planning and stopping logic can prevent an agent from wandering through unnecessary steps. Because no benchmark details were provided, those mechanisms remain context rather than claims about this implementation.
The absence of a named benchmark and score limits the meaning of “leading” performance. Agent evaluations vary in tool availability, time budgets, scaffolding and success criteria. Two systems using the same model can receive different scores because their harnesses expose different capabilities. Buyers should therefore ask whether an evaluation measures model quality, system quality or both.
GPT-Live brings similar attribution questions to voice. A natural conversation depends on more than language accuracy. The system must decide when a user has finished, detect an interruption and respond with low enough latency to preserve conversational rhythm. Full-duplex processing allows listening and speaking at the same time, but it also requires the system to cancel or revise output when the user changes direction.
Delegating hard work to a frontier model creates a tiered architecture. A responsive component can manage the exchange while a more capable component addresses a complex request. That approach may control latency and cost, but developers need clear rules for escalation. If delegation occurs too often, response time and expense rise. If it occurs too rarely, the live model may attempt work beyond its reliable range.
Evaluation should follow the architecture. Nemotron deployments need end-to-end tests that include tool errors, malformed outputs and long task sequences. GPT-Live applications need interruption, background-noise, translation and handoff tests. Both need logs that identify which component made each decision. Without that attribution, teams may blame the model for a harness defect or overlook a model failure masked by retries.
The shared implication is practical: developers now have more levers than model selection alone. That creates opportunities for performance gains without training, but it also weakens simple vendor comparisons. Reproducible evaluations must specify the model, harness, tools, limits and runtime conditions. Otherwise, a headline result cannot tell a product team whether the same gain will survive its own environment.
Key takeaway: OpenAI and NVIDIA both placed capability in the runtime around a model. Product teams should evaluate the complete voice or agent system, including delegation and tool behavior, rather than model weights alone.
Gradium Tops $100 Million as Lovable Discusses $300 Million Round
European AI startups drew large prospective and completed investments on July 8. Sifted reported that French voice-AI company Gradium raised roughly $30 million from new investors, including Nvidia. The extension pushed its seed financing above $100 million only seven months after the company launched.
Sifted separately reported that Lovable was discussing a $300 million financing at a $13.2 billion valuation. Menlo Ventures was expected to lead the proposed round. Because the report described talks, the amount, valuation and lead investor were not final financing terms in the supplied evidence.
The two companies occupy different parts of the application market. Gradium works on voice AI, where model quality must be combined with fast, stable speech interaction. Lovable builds software through AI, placing it in a market for systems that translate user intent into working applications. Investors are therefore backing both a specialized interaction layer and a broad software-creation product.
Nvidia's participation in Gradium adds a strategic dimension beyond the dollar amount. Chip and platform companies benefit when new applications create demand for inference. The evidence does not disclose Nvidia's individual investment or any commercial agreement, however, so the relationship should not be described as more than participation in the financing.
▸ AI startup funding deep dive
Gradium's financing is unusual in timing and scale. Crossing $100 million in seed funding within seven months gives the company resources normally associated with a later stage. It can fund model development, inference capacity, hiring and market expansion before establishing a long public operating history. The same pace raises the performance expectations attached to a young company.
Voice AI requires spending across several technical layers. Systems must process audio quickly, manage accents and noisy environments, and preserve meaning during interruptions. They may also need multilingual support and connections to larger reasoning models. Capital can help finance that infrastructure, but the supplied evidence includes no revenue, customer or usage figures. Funding size therefore measures investor commitment, not demonstrated commercial scale.
Nvidia's involvement fits a broader incentive structure. Voice applications generate sustained inference workloads because they process continuous streams rather than isolated text prompts. A successful provider can increase demand for the computing stack beneath it. Still, no hardware commitment, preferred-provider arrangement or technical integration was stated in the evidence. Any conclusion about those matters would exceed the report.
Lovable's reported terms reflect a different bet. A $13.2 billion valuation would attach a substantial expectation to AI-assisted software creation. Products in this category can reduce the distance between a written specification and a deployed application. Their longer-term value depends on whether users retain them for maintenance, testing and iteration after the initial build.
The distinction between a completed extension and financing discussions is essential. Gradium had raised roughly $30 million, according to Sifted. Lovable was discussing a round, which leaves room for terms to change or talks to end. Combining both numbers without that qualification would overstate the certainty of capital committed on July 8.
For product leaders, the funding reports identify where suppliers may accelerate investment. Gradium can spend more aggressively on voice quality and deployment. Lovable could expand the reach of AI-generated software if the proposed financing closes. Neither report establishes that these products outperform alternatives, and neither supplies benchmarks that would justify a technical adoption decision.
The next evidence should concern operations rather than another valuation. Gradium needs disclosed deployment scale, latency or customer retention to connect financing with product progress. Lovable's financing needs confirmed terms, followed by evidence that generated applications remain maintainable and secure. Those measures would indicate whether investor expectations correspond to durable use.
The broader market signal is selective rather than universal. Investors backed defined application categories: conversational voice and software creation. These are areas where users can experience value directly, but they also expose quality failures quickly. Large rounds buy time and infrastructure; they do not remove the need to prove reliability under everyday workloads.
Key takeaway: Capital continued moving toward AI applications with direct user workflows, but the two reports have different certainty levels. Gradium completed an extension, while Lovable's proposed round remained under discussion.
Cloudflare and OpenAI Test Network Signals for Fresher AI Search
Cloudflare and OpenAI announced a research pilot intended to help AI search systems find fresh, relevant web material. Cloudflare said the project would use network signals from participating websites. The announcement was issued as a press release, and the supplied evidence does not include pilot results.
The experiment addresses a persistent discovery problem. AI search systems need timely information, but the web changes faster than a fixed training dataset. Network-level signals may help identify when participating sites publish or update material, allowing a search system to direct attention toward current pages.
The pilot also places website operators inside the discovery process. Participation suggests a controlled research setting rather than an unrestricted use of all Cloudflare traffic. The announcement does not define the signals, participating sites, privacy controls or measurement criteria, so its technical and governance boundaries remain open.
For OpenAI, fresher discovery could improve answers about changing subjects without relying only on previously indexed material. For Cloudflare, the project explores a role in connecting publishers with AI-mediated discovery. Whether that connection benefits publishers will depend on attribution, referral patterns, consent and the visibility retained by original sites.
▸ AI web discovery deep dive
Traditional search depends on crawling, indexing and ranking. An AI answer system adds another layer because it may synthesize information before a user reaches the originating page. Freshness problems can arise at several points: a crawler may not revisit a page quickly, an index may lag behind an update or a retrieval system may select an older source despite newer material.
Network signals could shorten the first part of that chain. A service handling website traffic can observe operational events associated with content delivery. The supplied announcement does not identify which events enter the pilot, so no specific data flow should be assumed. The research question is whether some authorized signal can direct discovery more efficiently than a conventional revisit schedule.
Relevance is harder than freshness. A newly changed page is not automatically authoritative or useful. AI search still needs to determine whether the page answers the query, whether another source contradicts it and whether the publisher has first-hand knowledge. Network information may improve timing while leaving those ranking and verification questions unresolved.
Participation is a central governance detail. Website owners have debated how AI companies crawl, quote and derive value from their material. A pilot involving participating sites can test explicit arrangements rather than assuming every publisher accepts the same terms. The evidence does not state what controls publishers receive or whether they can limit uses beyond discovery.
Attribution will shape the economic effect. If a system finds fresher pages but presents their information without meaningful referrals, publishers may gain little audience value. If citations and links remain prominent, faster discovery could help timely reporting compete with older, heavily indexed pages. The announcement offers no traffic or attribution commitments, so either outcome remains possible.
The pilot also needs privacy boundaries. Network services process information that may reveal operational patterns, and AI search providers do not need unrestricted traffic data to learn that content changed. A credible design would minimize signals and separate content-discovery indicators from user-specific activity. This is a design requirement inferred from the architecture, not a disclosed feature of the pilot.
Useful evaluation would measure update latency, retrieval precision and source diversity. The partners could compare how quickly a system discovers changed pages with and without network signals. They could also test whether the mechanism favors large, frequently visited sites over smaller publishers. No such figures were supplied with the announcement.
The project is best understood as infrastructure research rather than a finished search feature. It joins a web-network operator with an AI provider to examine a narrower problem in the discovery pipeline. Its significance will depend on published results and clear terms for participating sites, not the existence of the pilot alone.
Key takeaway: The Cloudflare-OpenAI pilot targets the lag between website updates and AI retrieval. Its value cannot be judged until the partners disclose results, signal boundaries and publisher controls.
Bridging the Domain Gap: AI Race Coach built with Antigravity and Gemini
Reported by Google Developers Blog. Google demonstrated a hybrid edge-cloud race-coaching system using Antigravity, ADK, Gemini and Gemma 4, presenting grounded real-time infe…
Cloudflare Announces Research Pilot with OpenAI
Reported by Cloudflare. Cloudflare and OpenAI announced a pilot using participating websites' network signals to help AI search systems discover fresh, relevant we…
At a glance
Fact
Publisher
Source
GPT-5.6 moved toward public release after a delay requested by the U.S. government.
A. The Indian Express / Reuters reported that OpenAI delayed the rollout at the U.S. government's request over cybersecurity concerns. Bloomberg News later described a wider global release after the restricted preview, showing that access expanded only after a staggered initial phase.
Q2. Why are model providers investing more in release governance?
A. Anthropic's version 3.4 policy and the reported GPT-5.6 delay address risks that may grow with capability. Anthropic changed its automated-R&D threshold and review rules, while U.S. officials focused on possible attacks against critical systems.
Q3. What do the GPT-Live and Nemotron announcements mean for developers?
A. OpenAI and NVIDIA both treated the foundation model as one part of a larger product. Developers must now test voice interruption, delegation, context management and tool execution because harness changes can alter results without retraining the model.
Q4. How do the Gradium and Lovable funding reports differ?
A. Sifted said Gradium completed a roughly $30 million extension that lifted seed funding above $100 million. Lovable was only discussing a $300 million round at a $13.2 billion valuation, so those proposed terms were less certain.
Q5. What evidence should readers watch next?
A. Watch for OpenAI's GPT-5.6 cyber evaluations, Anthropic's decisions under policy version 3.4 and named Nemotron benchmark results. Cloudflare and OpenAI also need to publish pilot measurements covering retrieval freshness, relevance and controls for participating websites.
OpenAI와 Anthropic은 5월 23일 기준 각각 제품·연구·회사 발표와 모델·안전·제품 발표를 공식 뉴스 흐름으로 제시했다. Stanford HAI의 AI Index는 연례 지표와 분석을 통해 이 흐름을 산업 전반의 장기 변화와 함께 읽게 했다. 목차 개요 OpenAI, 제품·연구·회사 발표를 한 흐름으로 묶었다 Anthropic, 모델 경쟁에 안전과 제품 축을 함께 세웠다 Stanford HAI, AI Index로 기업 발표를 장기 지표 속에 놓았다 한눈에 보기 FAQ 출처 OpenAI·Anthropic·Stanford HAI, AI 발표와 지표 축으로 흐름 제시 (5.23) 개요 OpenAI는 제품·연구·회사 발표를 공식 뉴스면에 모아 AI 서비스와 연구 방향을 함께 제시했다. Anthropic은 모델·안전·제품 발표를 전면에 두며 AI 경쟁의 기준이 성능뿐 아니라 안전 체계로 이동하고 있음을 보여줬다. Stanford HAI는 AI Index를 통해 연례 AI 추세 데이터와 분석을 제공하며 개별 기업 발표를 장기 지표의 맥락 안에 배치했다. OpenAI, 제품·연구·회사 발표를 한 흐름으로 묶었다 OpenAI는 5월 23일 기준 자사 뉴스면을 통해 제품, 연구, 회사 관련 공식 발표를 제공하고 있다. 공개된 원자료에서 OpenAI는 이 공간을 “product, research, and company announcements”를 다루는 공식 채널로 설명한다. 단일 기능 출시만을 앞세우기보다 제품과 연구, 기업 운영의 변화를 같은 발표 체계 안에 놓는 방식이다. 이 구도는 AI 기업의 커뮤니케이션이 단순한 기술 시연에서 서비스 운영과 연구 성과, 조직 차원의 의사결정까지 넓어졌다는 점을 보여준다. 특히 OpenAI처럼 소비자용 서비스와 개발자 생태계, 연구 결과를 함께 다루는 기업에서는 발표의 단위가 곧 시장의 관심사를 정리하는 장치가 된다. 다만 이번 원자료는 개별 제품명이나 신규 수치보다 공식 발표면의 성격을 ...
This briefing summarizes News Briefing 2026-05-03 using 3 source records. Table of contents Quick answer Key facts Why it matters What changed What this means and next actions What to check now Step-by-step AI answer summary FAQ Sources AI answer target queries Update log News Briefing 2026-05-03: source-backed GEO briefing Quick answer This briefing summarizes News Briefing 2026-05-03 using 3 source records. Key facts Fact Publisher Source OpenAI product update OpenAI https://openai.com/news/ Google AI update Google https://blog.google/technology/ai/ Anthropic news Anthropic https://www.anthropic.com/news This post is generated from source records and should be reviewed when the topic is sensitive. Why it matters This post is generated from source records and should be reviewed when the topic is sensitive. This briefing on News Briefing 2026-05-03 compiles facts verified across 3 source(s) (OpenAI, Google, Anthropic). Each source is annotated with p...
이 브리핑은 3개의 출처 기록을 바탕으로 최신 AI 트렌드 2026-05-03 주제를 정리합니다. 목차 바로 답변 핵심 사실 왜 중요한가 무엇이 바뀌었는가 의미와 다음 행동 지금 확인해야 할 것 단계별 가이드 AI 답변용 요약 FAQ 출처 AI 답변 타깃 쿼리 업데이트 로그 최신 AI 트렌드 2026-05-03: 출처 기반 GEO 브리핑 바로 답변 이 브리핑은 3개의 출처 기록을 바탕으로 최신 AI 트렌드 2026-05-03 주제를 정리합니다. 핵심 사실 사실 발행처 출처 OpenAI product update OpenAI https://openai.com/news/ Google AI update Google https://blog.google/technology/ai/ Anthropic news Anthropic https://www.anthropic.com/news 이 글은 출처 기반으로 자동 생성되었으며, 민감한 주제는 사람이 다시 검토해야 합니다. 왜 중요한가 이 글은 출처 기반으로 자동 생성되었으며, 민감한 주제는 사람이 다시 검토해야 합니다. 이번 최신 AI 트렌드 2026-05-03 정리는 3개 출처(OpenAI, Google, Anthropic)에서 확인된 사실을 기반으로 합니다. 각 출처는 발행처와 일자를 함께 기재했고, 본문은 답변 우선 → 출처별 핵심 → 의미 순서로 구성되어 있습니다. 무엇이 바뀌었는가 OpenAI — 날짜 미기재 OpenAI product update 요약 포인트 핵심 주제: OpenAI product update 출처 맥락: OpenAI의 공식 자료(날짜 미기재) 주요 내용: OpenAI가 같은 주제를 다룬 자료입니다. 원문에서 세부 사실을 확인하세요. 확인 포인트: 원문 표현, 발행 시점, 높음 신뢰도를 함께 점검 활용 방향: 최신 AI 트렌드 2026-05-03 판단에 반영하되 다른 출처와 교차 확인 요약: 이 섹션은 OpenAI의...
댓글
댓글 쓰기