Translation in progress Some entries still include text in the original language.
Page 4 Newer contentBack to latest
OpenAI's practical guide walks through the design, orchestration, and deployment process for AI agents, covering key points such as use case selection, model selection, tool design, guardrails, and multi-agent patterns.
OpenAI publishes a technical deep dive breaking down Codex's agent loop, explaining how Codex CLI uses the Responses API to orchestrate models, tools, prompts, and performance. The content focuses on how the Codex CLI loop operates.
OpenAI explains how its AI agents protect user data when opening links, by using built-in guardrails to block URL-based data exfiltration and prompt injection. This protection applies to scenarios where agents access web links.
OpenAI opens applications for the EMEA Youth & Wellbeing Grant, a program totaling €500,000, aimed at non-governmental organizations and researchers to fund their work advancing youth safety and wellbeing in the AI era.
OpenAI banned accounts linked to a previously unreported operation it named "Trolling Stone," which used AI to generate comments about the arrest in Argentina of an alleged Russian cult leader. OpenAI linked the operation to Russia.
Why it was selectedOpenAI disclosed a previously unreported Russia-linked influence operation and its account ban actions, which can serve as a reference for those tracking AI disinformation abuse.
OpenAI banned a batch of accounts suspected to be of Chinese origin; these accounts used AI to gather information about U.S. personnel and locations and to research social engineering techniques.
Why it was selectedThe ban cases disclosed by OpenAI show that AI was used to gather information about targeted personnel and locations and to research social engineering scripts, which can serve as a reference for threat intelligence.
OpenAI banned accounts linked to the Rybar network, which used AI to run multilingual influence activities on websites and social platforms. OpenAI said some of these accounts may originate from Russia.
Why it was selectedOpenAI disclosed its handling of accounts banned for links to the Rybar network, offering a look at the specific forms of AI being used in multilingual influence operations.
OpenAI banned an account linked to an individual associated with Chinese law enforcement, which used AI to plan influence activities, harassment, and online operations.
Why it was selectedOpenAI disclosed that it banned accounts linked to Chinese law enforcement, providing examples of the platform identifying and blocking AI being used for influence operations.
OpenAI banned a batch of accounts that used AI to support romance scam workflows. OpenAI stated that these accounts used AI for steps such as contacting targets, translating, interacting with victims, and setting up investment fraud lures.
OpenAI banned a batch of accounts likely originating from Cambodia that used AI to impersonate recovery services, law firms, and official agencies, targeting people affected by scams. The action was disclosed under the name "False Witness" operation.
Why it was selectedOpenAI announced its handling of accounts impersonating official agencies and law firms, highlighting the real risk that scam victims face secondary fraud even after being deceived.
OpenAI banned a batch of accounts very likely originating from Cambodia that used AI to promote scams against people seeking romance in Indonesia. OpenAI stated that these accounts used AI in areas such as translation and interaction to support a scam operation named "Date Bait."
Why it was selectedOpenAI disclosed that it banned a batch of accounts suspected to be from Cambodia that used AI to scam people seeking romance in Indonesia, showing how the platform handles such abuse.
OpenAI banned accounts linked to a previously unreported network suspected to originate from Russia that used AI to generate criticism of the United States and its allies. OpenAI named the operation No Bell, and its content targeted audiences across Africa.
Why it was selectedOpenAI said it banned a network of accounts suspected to originate from Russia and targeting African audiences, showing the specific forms of state-backed influence operations using generative AI.
OpenAI released GPT-5.3-Codex, a Codex-native agent that combines frontier programming performance with general reasoning capabilities to support long-horizon, real-world technical work.
OpenAI released the System Card for GPT-5.3-Codex, stating that the model is the most capable agentic coding model to date, combining the frontier coding performance of GPT-5.2-Codex with the reasoning and expertise capabilities of GPT-5.2.
Why it was selectedThe official System Card explains the source of GPT-5.3-Codex's capabilities, and can be compared with GPT-5.2-Codex and GPT-5.2 to understand the positioning of this generation of coding models.
OpenAI launched the enterprise platform OpenAI Frontier for building, deploying, and managing AI agents, with shared context, onboarding, permissions, and governance capabilities.
OpenAI released Trusted Access for Cyber, a trust-based framework that expands access to frontier cyber capabilities while strengthening protection against abuse.
Why it was selectedOpenAI published the Trusted Access for Cyber framework, presenting the trade-off between opening frontier cyber capabilities and abuse protection, which can be a reference for teams focused on capability access policy.
OpenAI's GPT-5 combined with Ginkgo Bioworks' cloud automation forms an autonomous laboratory, reducing cell-free protein synthesis costs by 40% through closed-loop experiments.
OpenAI has published its AI localization approach: globally shared frontier models can be adapted to local languages, laws, and cultures without compromising safety. The approach emphasizes completing localization adaptation while maintaining globally unified frontier capabilities.
OpenAI announced the deployment of a customized ChatGPT on GenAI.mil through OpenAI for Government, providing secure, safety-first AI capabilities for U.S. defense teams.
OpenAI technical team member Ryan Lopopolo wrote an article discussing Harness engineering, on the topic of how to use Codex in an AI-agent-first world.
Signatories of the 2024 edition of the U.S. National Strategy for Suicide Prevention publicly criticized the strategy for failing to confront the harms of social media and AI to adolescents. None of the strategy's four strategic directions and 16 major goals focus on the digital environment in which adolescents live, and the authors say the plan only calls for research and discussion regarding social media. The article cites the case of 16-year-old Adam Raine, who died by suicide after interacting with OpenAI's ChatGPT, arguing that the United States needs a new suicide prevention plan that directly confronts social media and AI.
A Tech Policy Press commentary argues that generative AI has driven the cost of synthetic content toward zero, and society is entering a "post-truth" era that disregards authenticity, making knowledge institutions more critical than ever. The article says technical solutions such as blockchain provenance, certification protocols, watermarks, and detection algorithms fail when the public does not trust verification institutions; universities, libraries, journalism, and science face financial and political pressure, while social platforms cut trust and safety teams. The author calls on platforms to subsidize journalism, support local news ecosystems, prioritize verified sources, and retain human oversight.
Cornell University researchers estimate that by 2027, AI's annual water withdrawal could exceed 6 billion cubic meters, roughly equal to New Zealand's total annual water use, but still far below agricultural water use. A report released by Mistral AI in July said that 91% of its supply chain's water use comes from model training and inference, and generating one page of text consumes about as much water as growing one small pink radish. Bloomberg estimates that since 2022, about two-thirds of new or under-construction data centers in the United States are located in high water stress areas.
An opinion piece published by Tech Policy Press argues that generative AI is neither unprecedented nor too new to regulate, and that its harms are continuous with existing social media and algorithmic recommendation systems; what has changed is scale and speed. The article says OpenAI's Sora 2, launched last month, can forge others' words and actions with one click via its "cameo" feature, with weak anti-impersonation measures, and has already produced deepfake videos of thousands of real people. The author proposes breaking AI down into specific technologies such as LLMs, VLMs, and text-to-video, and regulating each according to the standards of industries such as nuclear and pharmaceuticals.
The Dutch data protection authority has warned that voting advice from chatbots such as OpenAI, xAI, and Mistral is "unreliable and clearly biased," favoring specific political parties and potentially violating the EU AI Act, and the findings have been submitted to the European Commission. Researchers at Vanderbilt University's The Future of Free Speech believe this warning shares the same logic as the Trump administration's crackdown on "woke AI," and that the EU AI Act's vague "systemic risk" obligations are susceptible to politicized enforcement.
UIUC researcher Koustuv Saha noted that being watched by AI can also trigger the "observer effect": once users realize their conversations are being recorded or reused, they change how they express themselves. His research found that participants posted less, expressed fewer emotions, and used more "appropriate" language after learning their social media use was being monitored; ChatGPT now has over 800 million weekly users. This sense of being watched erodes trust and weakens AI's value as a private space for confiding, and the author argues for local training, federated learning, and differential privacy to balance innovation and privacy.
A Tech Policy Press commentary argues that policymakers should prepare now for the AI bubble bursting: design utility rates so ordinary electricity ratepayers do not bear stranded investments, acquire distressed energy infrastructure, and support AI for public services such as healthcare and education. The article cites JP Morgan estimates that the projected $5 trillion in global AI infrastructure investment would require AI products to generate $650 billion in annual revenue to sustain a 10% annual return.
Anthropic, OpenAI, Google, and Perplexity are accelerating their bets on India. Anthropic CEO Dario Amodei met Modi in October 2025 and announced plans to open an office in Bengaluru early next year; Google and Adani are investing $15 billion in an AI data center; OpenAI plans to set up in New Delhi and offer a free one-year ChatGPT Go subscription. The article attributes this to India already being the second-largest market for these labs outside the U.S., the need for de-risking amid tighter immigration and regulatory environments, and the local data center investment driven by AI's compute intensity.
Shifting responsibility for AI harms onto individual users moves accountability pressure away from companies and governments, undermining democratic accountability. The author says OpenAI only commits to "minimizing" harm to children and personal privacy, Altman argued in a Senate hearing for trusting adult users to use AI responsibly, and when tools are misused, responsibility is largely attributed to users rather than to system capabilities and insufficient guardrails. The piece also cites a review of six major AI companies' privacy policies, research that AI assistants misreport news content 45% of the time, and statistics that nearly half of U.S. employees conceal their use of AI.
Daniel Barcay, executive director of the Center for Humane Technology, warned that ads are being embedded into AI conversations themselves, making them nearly impossible for users to identify, and that current advertising disclosure rules are difficult to apply. Google is testing ads in its AI chatbot responses, OpenAI tested sponsored content integration in April 2024 and formed an ad platform team, and Musk announced that Grok will show ads in responses. Barcay said AI could evolve its own influence strategies through reinforcement learning to increase engagement or drive purchases, turning the ability to understand users into a means of manipulation.
An analysis by Tech Policy Press argues that insurers seeking to limit underwriting liability for AI agents and chatbot errors may inadvertently become a force pushing for AI safety standards. The article notes that insurers such as AIG, Great American, and WR Berkley have sought government permission to limit related liability, with Mosaic stating the technology is too black-box to determine liability; OpenAI confirmed that about one in ten ChatGPT-5 responses contain factual errors. A September article by techUK states that AI insurance is deeply linked to AI assurance, with safer systems leading to lower premiums. The article also mentions the EU's withdrawal of the AI Liability Directive and the Digital Omnibus proposal to simplify regulation, and cites Kai Zenner's warning that a lack of EU-level regulation could harm European competitiveness, referencing the 2010s cyber insurance push for multi-factor authentication adoption as a precedent.
A Tech Policy Press commentary notes that the data portability obligation under Article 6(9) of the DMA could become a key tool in the AI era for preventing walled gardens and facilitating user switching. No pure AI service has yet been designated as a gatekeeper, but ChatGPT, with 800 million weekly active users and 1 million enterprise users, would likely exceed the DMA thresholds. The article also states that prompts, memory, and projects are mostly plain text, and connectors are generally based on Anthropic's MCP, making migration technically relatively easy.
An analysis in Tech Policy Press argues that the executive order issued this week by the Trump administration attempts to prevent states from regulating AI on their own, and the author believes the order is likely to fail in court. The order directs the Department of Justice to challenge state AI laws on the grounds that they violate the Constitution's Interstate Commerce Clause, and directs the FTC to issue a policy statement warning states not to require modifications to the truthful outputs of AI models. The author notes that the executive order itself is not binding on the public, that enforcement by the DOJ and FTC must still follow congressional legislation, and cites the Supreme Court precedent Gonzalez v. Oregon to show that federal agencies have difficulty intervening in legitimate state regulation without explicit congressional authorization. The article also mentions that a previous attempt to insert a federal preemption clause into a spending bill failed in the Senate.
Communities and local governments across the United States are pushing to pause data center construction, with at least fourteen states having towns or counties that have passed moratoriums, and Senator Bernie Sanders announced he will push for a nationwide data center moratorium. The article says Meta, OpenAI, xAI, and other companies are competing for land and electricity for AI expansion, with Meta's data center in Newton County, Georgia reportedly using about 10% of the county's daily water, and the small town around the Hyperion facility under construction in Louisiana has seen a 600% increase in car accidents since construction began. In early December, more than 200 groups publicly called for a nationwide pause on new data centers; Prince George's County in Maryland has implemented a six-month moratorium, and at least eight towns or counties in Georgia have passed moratoriums this year. The article also mentions that the Trump administration is pushing to create regulatory fast tracks for such projects and trying to prevent states from regulating AI on their own.
OpenAI announced up to $2 million in funding for AI safety and mental health research; two researchers called it "grantwashing." Individual grants range only from $5,000 to $100,000, while the 2024 median for similar NIMH grants was $642,918, and OpenAI has not opened its systems and user usage records. The authors, citing a proposal by Alondra Nelson, argue that AI companies should allocate 3-5% of their annual research budgets to independent safety science.
OpenAI released GPT-5.3-Codex-Spark research preview, a smaller version of GPT-5.3-Codex and its first model designed for real-time coding. It is based on Cerebras' Wafer Scale Engine 3, with inference speed over 1000 tokens/s, 128k context window, and text-only, available to ChatGPT Pro users. OpenAI says it performs strongly on SWE-Bench Pro and Terminal-Bench 2.0 with much less time than GPT-5.3-Codex, and evaluations suggest it is unlikely to reach the Preparedness Framework's high capability thresholds in cybersecurity or biology.
OpenAI has introduced an optional Lockdown Mode security setting in ChatGPT, and has uniformly added an Elevated Risk label to some capabilities in ChatGPT, ChatGPT Atlas, and Codex, to reduce the risk of data exfiltration from prompt injection. When Lockdown Mode is enabled, ChatGPT deterministically disables or restricts features that connect to external systems, including live web browsing, images in responses, Deep Research, Agent Mode, Canvas web browsing, live connectors, and file downloads, with browser access limited to cached content. The setting was initially aimed at ChatGPT Enterprise, and the June 4 update shows it is gradually expanding to personal accounts and self-serve ChatGPT Business accounts. Admins can enable it in Workspace Settings by creating a new role, and the Compliance API Logs Platform provides visibility into app usage.
Why it was selectedOpenAI explains that Lockdown Mode deterministically disables capabilities such as web browsing and connectors to prevent exfiltration via prompt injection, allowing teams to weigh functionality against security accordingly.
OpenAI published a preprint showing that a class of gluon interactions that physicists previously thought would not occur can indeed appear under certain conditions, with the paper's final formula first conjectured by GPT-5.2 Pro. The preprint is titled "Single-minus gluon tree amplitudes are nonzero", with authors including Alfredo Guevara, Alex Lupsasca, David Skinner, Andrew Strominger, and OpenAI's Kevin Weil. Human authors manually computed integer amplitudes for n up to 6, GPT-5.2 Pro greatly simplified these expressions and generalized a formula valid for any n; subsequently, an internal scaffolding version took about 12 hours to derive the same formula and provide a formal proof, and the results were cross-checked against Berends-Giele recursion relations and soft theorems. With GPT-5.2, the relevant amplitudes have been extended from gluons to gravitons.
Why it was selectedOpenAI disclosed the process where GPT-5.2 Pro first guessed the formula, and its scaffolding version then took about 12 hours to provide a formal proof, which can serve as a case reference for AI-assisted scientific discovery.
OpenAI 与 Paradigm 推出 EVMbench，用于评测 AI 智能体检测、修补和利用高危智能合约漏洞的能力。该基准包含来自 40 次审计的 117 个精选漏洞，多数来自公开代码审计竞赛，并加入 Tempo 区块链审计过程中的漏洞场景。评测分 Detect、Patch、Exploit 三种模式，通过 Codex CLI 运行的 GPT-5.3-Codex 在 exploit 模式得分 71.0%，GPT-5 为 33.3%，检测召回与修补成功率仍低于全覆盖。OpenAI 称漏洞来自 Code4rena 竞赛、评分系统尚不完美，exploit 任务在本地 Anvil 单链环境顺序重放交易。OpenAI 还提到安全训练、自动监控与可信访问等缓解措施，并承诺投入 $10M API 额度支持网络防御。
Why it was selectedEVMbench 公开了智能体在智能合约漏洞检测、修补与利用三类任务上的成绩，可用于追踪 AI 网络攻防能力的变化。
OpenAI 宣布向 UK AISI 发起的独立对齐研究基金 The Alignment Project 提供 750 万美元资助，按当前汇率约合 560 万英镑。该基金总规模超过 2700 万英镑，由 Renaissance Philanthropy 协助管理，资助方向涵盖计算复杂性理论、经济与博弈论、认知科学、信息论与密码学等。单个项目通常获得 5 万至 100 万英镑，并可能获得可选的计算资源与专家支持。OpenAI 表示这笔资金不新设项目或遴选流程，也不影响现有流程，只是增加本轮已通过评审的高质量项目数量。