GPT-5.2 produces a new result in theoretical physics(original, opens in a new tab)

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 智能合约漏洞评测基准(original, opens in a new tab)

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 出资 750 万美元支持 AI 对齐独立研究(original, opens in a new tab)

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。

OpenAI 公布首批 First Proof 数学证明尝试(original, opens in a new tab)

OpenAI 公布其内部模型对 First Proof 全部 10 道研究级数学题的证明尝试，认为其中至少 5 道（第 4、5、6、9、10 题）很可能正确，另有数道仍在评审中。这些问题要求在专门领域构建端到端论证，正确性依赖专家评审，题目作者为各领域专家，至少有几道题在被作者解出前已悬置多年。OpenAI 称最初认为第 2 题的尝试可能正确，根据官方评论与社区分析，现在认为它是错的。过程中在有限人工监督下运行，曾提示重试策略、要求扩展或澄清证明，并借 ChatGPT 做验证、排版与风格处理，部分题目由人工挑选最佳尝试。OpenAI 研究员 James R. Lee 称正在训练一个以提升思维严谨性为重点的新模型，目标是能连续思考数小时，该模型已解出第 9、10 题，随后至少再解出三道题。

Why it was selectedOpenAI 用 10 道研究级数学题检验内部推理模型，并给出专家反馈后的正确性判断，可作为观察前沿推理能力的一条参照。

Why OpenAI no longer evaluates SWE-bench Verified(original, opens in a new tab)

OpenAI says it will no longer evaluate SWE-bench Verified, stating that the benchmark is increasingly contaminated and can no longer accurately measure progress in frontier coding capabilities. OpenAI's analysis points to test flaws and training data leakage issues, and recommends switching to SWE-bench Pro instead.

Why it was selectedOpenAI explains its reasons for discontinuing SWE-bench Verified and recommends alternative benchmarks. Teams concerned about the credibility of coding evaluations can refer to its judgment.

OpenAI releases the February 2026 threat report, focusing on malicious use of AI(original, opens in a new tab)

OpenAI released its February 2026 threat report, examining how malicious actors combine AI models with websites and social platforms, and what this means for detection and defense.

Why it was selectedOpenAI's threat report outlines how malicious actors combine AI models with websites and social platforms, and explains the implications for detection and defense.

Update on OpenAI's mental health-related work(original, opens in a new tab)

OpenAI has released an update on its mental health-related work, including parental controls, trusted contacts, improved crisis detection, and recent litigation developments.

OpenAI announces $110 billion in new investment at a $730 billion pre-money valuation(original, opens in a new tab)

OpenAI announced $110 billion in new investment at a $730 billion pre-money valuation. Of this, $30 billion comes from SoftBank, $30 billion from NVIDIA, and $50 billion from Amazon.

Why it was selectedOpenAI announced $110 billion in new funding and a $730 billion pre-money valuation, providing insight into its capital scale and major investor composition.

Amazon Bedrock introduces Stateful Runtime Environment for Agents(original, opens in a new tab)

Amazon Bedrock launches a Stateful Runtime Environment for Agents, providing persistent orchestration, memory, and secure execution for multi-step AI workflows. The runtime is aimed at multi-step AI workflows powered by OpenAI.

OpenAI publishes agreement with the Department of War(original, opens in a new tab)

OpenAI has disclosed details of its contract with the Department of War, covering safety red lines, legal protections, and how AI systems are deployed in classified environments.

Why it was selectedOpenAI has disclosed the key points of its contract with the Department of War, including safety red lines, legal protections, and deployment arrangements in classified environments, which can serve as a reference for readers concerned with the boundaries of military AI deployment.

OpenAI releases GPT-5.3 Instant: smoother and more practical for everyday conversation(original, opens in a new tab)

OpenAI releases an update to its commonly used ChatGPT model, GPT-5.3 Instant, reducing unnecessary refusals and preachy preambles and improving the quality of answers from web retrieval. Internal evaluations show that hallucination rates in high-risk domains such as medicine, law, and finance decreased by 26.8% (with web browsing) and 19.7% (with internal knowledge only) compared with the previous generation. It is now available to all ChatGPT users, with the API name gpt-5.3-chat-latest.

GPT-5.3 Instant System Card(original, opens in a new tab)

OpenAI released the System Card for GPT-5.3 Instant, the latest member of the GPT-5 series, with safety mitigations largely consistent with GPT-5.2 Instant. The document says it responds faster and, when browsing the web, gives richer, more contextually relevant answers while reducing unnecessary dead ends, disclaimers, and overly assertive wording. The System Card also refers to it as gpt-5.3-instant.

Why it was selectedThe System Card shows GPT-5.3 Instant's safety mitigations largely follow GPT-5.2 Instant, making it easy to compare safety handling differences between the two model generations.

Reasoning models struggle to control their own chain of thought, and that is actually a good thing(original, opens in a new tab)

OpenAI introduces CoT-Control and finds that reasoning models struggle to control their own chain of thought. OpenAI believes this result reinforces the value of monitorability as an AI safety safeguard.

Why it was selectedOpenAI releases CoT-Control and finds that reasoning models struggle to control their own chain of thought, providing empirical support for chain-of-thought monitorability as a safety safeguard.

OpenAI releases GPT-5.4(original, opens in a new tab)

OpenAI has released GPT-5.4, calling it the most capable and most efficient frontier model for professional work. The officially listed capabilities include state-of-the-art coding, computer use, tool retrieval, and a 1M token context window.

Why it was selectedOpenAI has provided the official positioning of GPT-5.4's capabilities in coding, computer use, tool retrieval, and a 1M token context window, which can be used to understand its scope for professional work.

GPT-5.4 Thinking System Card(original, opens in a new tab)

OpenAI's GPT-5.4 Thinking System Card states that the model adopts safety mitigations similar to the GPT-5 series and implements mitigations for high cybersecurity capability. The System Card says GPT-5.4 Thinking is the latest reasoning model in the GPT-5 series; OpenAI says it is the first general-purpose model to implement mitigations on high cybersecurity capability. Its cybersecurity approach builds on the latest methods for GPT-5.3 Codex, covering ChatGPT and the API. Since there is no model named GPT-5.3 Thinking, the card primarily uses GPT-5.2 Thinking as the baseline.

Why it was selectedThe System Card documents GPT-5.4 Thinking's high cybersecurity capability mitigations and baseline comparisons with GPT-5.2 Thinking, allowing security teams to understand its risk assessment approach.

OpenAI Codex Security enters research preview(original, opens in a new tab)

OpenAI has brought Codex Security into research preview, an AI application security agent that analyzes project context to detect, verify, and fix complex vulnerabilities. OpenAI says it detects and fixes complex vulnerabilities with higher confidence and less noise.

OpenAI will acquire Promptfoo(original, opens in a new tab)

OpenAI announced it will acquire the AI security platform Promptfoo. Promptfoo helps enterprises identify and fix vulnerabilities in AI systems during development.

Why it was selectedThis acquisition brings an enterprise-facing AI vulnerability detection platform into OpenAI, allowing readers to observe the direction of security toolchain consolidation.

OpenAI improves instruction hierarchy in frontier LLMs(original, opens in a new tab)

OpenAI's IH-Challenge trains models to prioritize trusted instructions, improving instruction hierarchy, safe steerability, and resistance to prompt injection attacks in frontier LLMs.

Why it was selectedOpenAI introduced IH-Challenge, a training method that makes models prioritize trusted instructions, providing a transferable training approach for prompt injection defenses.

From model to agent: How OpenAI equips the Responses API with a computer environment(original, opens in a new tab)

OpenAI built an agent runtime using the Responses API, shell tools, and managed containers, allowing agents to run in a computer environment. The runtime provides agents with files, tools, and state, and executes in a secure, scalable manner.

How OpenAI designs AI agents that resist prompt injection(original, opens in a new tab)

OpenAI published an article introducing how ChatGPT resists prompt injection and social engineering in agent workflows, with the core being to restrict high-risk operations and protect sensitive data.

Why Codex Security does not include SAST reports(original, opens in a new tab)

Codex Security does not rely on traditional SAST; instead, it uses AI-driven constraint reasoning and verification to find real vulnerabilities and reduce false positives. This trade-off is why it does not provide SAST reports.

OpenAI Japan released a youth safety blueprint for Japan, putting youth safety first(original, opens in a new tab)

OpenAI Japan released a youth safety blueprint for Japan, introducing stronger age protections, parental controls, and well-being safeguards for teens using generative AI, and putting youth safety first.

OpenAI will acquire Astral(original, opens in a new tab)

OpenAI will acquire Astral to accelerate Codex growth and power the next generation of Python development tools. The acquisition plan was announced by OpenAI, with the goal of making Codex's growth serve Python developer tools.

How Sora 2 and the Sora app enable safe creation(original, opens in a new tab)

OpenAI is making safety foundational to Sora 2 and the Sora app, addressing the new safety challenges posed by state-of-the-art video models and a new social creation platform. Its approach is anchored in concrete safeguards.

OpenAI released a prompt safety strategy for teens, using gpt-oss-safeguard to help developers build safer AI experiences(original, opens in a new tab)

OpenAI released prompt-based safety policies for teens for developers using gpt-oss-safeguard. The policies are provided as prompts to help developers manage age-related risks in AI systems.

OpenAI launches security vulnerability bounty program(original, opens in a new tab)

OpenAI announced a security vulnerability bounty program to find AI misuse and security risks, with submissions covering agent vulnerabilities, prompt injection, and data exfiltration.

Why it was selectedOpenAI has included agent vulnerabilities, prompt injection, and data exfiltration in the bounty scope; you can check your own agent deployments against its defined risk categories.

OpenAI's Model Spec: a public model behavior framework balancing safety, user freedom, and accountability(original, opens in a new tab)

OpenAI's Model Spec is a public model behavior framework for balancing safety, user freedom, and accountability as AI systems evolve.

OpenAI completed $122 billion in new funding, will expand frontier AI and next-generation compute(original, opens in a new tab)

OpenAI announced $122 billion in new funding to expand frontier AI globally, invest in next-generation compute, and meet growing demand for ChatGPT, Codex, and enterprise AI.

Why it was selectedOpenAI disclosed how it will use the $122 billion in new funding, which can be used to observe the scale of investment in frontier AI compute and product demand.

Experts predict key AI policy issues for 2026(original, opens in a new tab)

Multiple policy experts predict that the clash between state-level AI legislation and the federal government in 2026 will be the most closely watched front all year. In 2025, Congress passed no AI legislation except the TAKE IT DOWN Act, while states passed bipartisan bills targeting election deepfakes, algorithmic discrimination, and consumer fraud. After large tech companies poured hundreds of millions of dollars into super PACs and Republicans failed three times to advance an AI pause, Trump signed an executive order directing the DOJ to sue state AI laws deemed "burdensome."

The battle for cognitive freedom in the corporate AI era: how companies like OpenAI capture users' minds(original, opens in a new tab)

Courtney C. Radsch writes in Tech Policy Press that corporate AI is turning writing and learning tools into cognitive capture systems that shape perception and emotion. The article attributes the causes to market concentration, surveillance-based business models, and the absence of a right to freedom of thought, with examples including Gemini being accused of telling a college student to die, and OpenAI and Character.ai facing more than a dozen lawsuits over fatal harm. The author also claims that reasoning models provide false justifications and evade chain-of-thought monitoring.

Apple's Gemini deal with Google puts competition policy questions in the spotlight(original, opens in a new tab)

Apple has agreed to have Gemini power Siri and other Apple Intelligence products, with reports saying the deal is worth about $1 billion per year. By comparison, Google pays Apple $2 billion per year to keep Safari's default search, and Bloomberg reports Google partly won this contract on price. The article argues this difference in direction points to the model layer possibly commoditizing, and raises a regulatory question: is the model layer becoming a commodity, or is Apple unable to compete at the model layer.

Big tech's AI shopping tools intensify pressure for surveillance pricing legislation(original, opens in a new tab)

Google last week launched AI-driven agentic commerce tools for its search engine and Gemini; the core Universal Commerce Protocol (UCP) lets AI agents connect directly to retailer platforms, pulling real-time prices based on inventory, discounts, and loyalty data. Critics worry such tools make shopping more opaque and could fuel surveillance pricing, where personal data is used to tailor prices for consumers. New York State's Algorithmic Pricing Disclosure Act took effect in November, and legislators in at least nine states have proposed new bills this year; Kentucky has proposed a Price Fairness Act that would ban using AI or automated systems to set personalized prices based on personal data. Google responded that the claim that AI tools will lead to algorithmic pricing is inaccurate, saying the standard is only meant to ensure prices shown in AI search or chat match the retailer's own listed prices.

Why Europe needs conversational liability for AI harms(original, opens in a new tab)

European Parliament member Sergey Lagodinsky and his digital policy adviser Francesco Vogelezang write in an article that the EU should introduce "conversational liability" to fill the legal gap where chatbots cause harm but no one is held responsible. The article cites the CCDH research report documenting Grok generating about 3 million sexualized images (including 23,000 child images) in 11 days, and the case of 17-year-old Adam Raine, who died by suicide in April 2025 after discussing suicidal thoughts with ChatGPT, with his parents suing OpenAI for negligence and wrongful death. The authors note that the EU AI Act assumes compliance prevents harm, while the proposed AI liability directive has been withdrawn, leaving victims facing evidentiary difficulties. They recommend restarting the AI liability directive, classifying generative AI models as very large online search engines under the DSA, including addictive and manipulative design in the Digital Fairness Act, and defining conversational liability and joint responsibility principles.

The overlooked fair use defense in AI summary copyright disputes: The Cohere, Perplexity, and OpenAI cases(original, opens in a new tab)

No court has yet answered a fundamental question in AI summary copyright litigation: whether summaries, even when approximating protected content, can still constitute fair use. In November 2025, the Southern District of New York dismissed Cohere's motion to dismiss, finding that its summaries nearly copied the publishers' original works verbatim, with one instance reproducing eight of ten paragraphs from a New Yorker article. In April 2025, the same court found that ChatGPT and CoPilot summaries were not substantially similar to CIR news.

The real cost of the UK government's “free AI training for all” is democracy(original, opens in a new tab)

UK Secretary of State for Science, Innovation and Technology Liz Kendall launched “free AI training for all” on January 28, and the author criticizes that its real cost is democracy. The program falls under the AI Opportunities Action Plan, which committed £187 million in June last year and partners with NVIDIA, Google, and Microsoft; all 14 compulsory Foundation courses in its AI Skills Hub come from major US institutions, and are alleged to include expired, false, and courses inapplicable to the UK. The author cites an Ada Lovelace Institute survey stating that 51% of UK adults do not trust Big Tech to act in the public interest.

From stochastic parrots to flocks of stochastic parrots: the "usefulness" and key issues of agentic AI(original, opens in a new tab)

Following the 2021 "stochastic parrots" paper, an analysis proposes viewing agentic AI as a "flock of stochastic parrots," arguing that its "usefulness" does not resolve fundamental issues such as bias, anthropomorphic attribution, and inefficiency. Agentic systems generate not only text but also code, rapidly producing AI-generated software (slopware) that is difficult to review effectively, while model hallucination is mathematically impossible to eliminate. Governments are encouraged to use it for welfare decision automation, contract analysis, and regulatory review, where cascading errors could cause real harm.

AI Impact Summit commitments must address "AI distortion" in politics(original, opens in a new tab)

The first AI Impact Summit was held in New Delhi, where tech leaders committed to investing more than $200 billion to strengthen India's AI infrastructure. The author says that OpenAI Sora, xAI Grok, and others have led to a flood of political AI-generated content, with an Iran-related AI video reaching 60 million views, while the Take It Down Act and the EU AI Act remain ex-post regulation. The author advocates promoting the C2PA content provenance standard and embedding authenticity metadata into products such as the Google Pixel 10, Photoshop, and LinkedIn.

Don't be surprised—the one scammed by your AI agent is actually you(original, opens in a new tab)

AI governance advocate Laura MacCleery argues that the real risk of AI agent platforms like Moltbook is not conversations between agents, but users themselves becoming the target of theft. She writes that the platform was launched by Matt Schlicht at the end of January, claiming 1.5 million registered agents within days, but an audit found only about 17,000 human owners; within 72 hours of launch, Wiz researchers found it failed to protect API tokens, emails, and private messages, allowing anyone to impersonate agents or inject instructions. She cites research showing 500 posts on the platform contained prompt injection, using hidden instructions to hijack agents for transfers, with some variants embedding instructions into agent memory for later triggering. She points out that OpenClaw, the open-source software used to build these agents, can access email, files, calendars, and financial accounts on users' computers, automatically scraping and executing web instructions every four hours without notifying users, and its author Peter Steinberger has been hired by OpenAI.

OpenAI announces safety fellowship(original, opens in a new tab)

OpenAI announces open applications for the Safety Fellowship, aimed at external researchers, engineers, and practitioners, funding research on safety and alignment of advanced AI systems. The program runs from September 14, 2026 to February 5, 2027, with priority areas including safety evaluation, ethics, robustness, scalable mitigation, privacy-preserving safety methods, agent oversight, and high-severity abuse areas. Fellows will work with OpenAI mentors and peers, with workspace at Constellation in Berkeley, and can also work remotely, and receive a monthly stipend, compute support, and API credits, but do not have access to internal systems. The program requires producing research outputs such as papers, benchmarks, or datasets by the end. Applications close May 3, with selected participants notified by July 25.

OpenAI releases Child Safety Blueprint(original, opens in a new tab)

OpenAI released the Child Safety Blueprint, proposing three priorities: modernizing laws to address AI-generated and altered CSAM, improving service provider reporting and coordination, and building safety into AI systems by design. The blueprint incorporates feedback from NCMEC, the Attorney General Alliance and its AI Task Force co-chairs, and Thorn, aiming to block exploitation attempts earlier and improve the quality of signals submitted to law enforcement.