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.
OpenAI responded to the supply chain attack on Axios developer tools, rotating the macOS code signing certificate and updating the affected applications, while confirming that no user data was leaked.
OpenAI expanded the Trusted Access for Cyber program, introducing GPT-5.4-Cyber to vetted defenders while simultaneously strengthening safety guardrails. OpenAI stated that this change comes amid continued advances in AI cybersecurity capabilities, and the program is open to vetted defensive parties.
OpenAI updated the Agents SDK, adding native sandbox execution and a model-native harness to help developers build secure, long-running agents that can operate across files and tools.
OpenAI announced that several leading security companies and large enterprises have joined its Trusted Access for Cyber program, using GPT-5.4-Cyber with a $10 million API credit to strengthen global cyber defense. The program combines model capabilities with API funding support for participating organizations to use in cyber defense efforts.
OpenAI launched GPT-Rosalind, a reasoning model for life sciences research, designed to accelerate drug discovery, genomics analysis, protein reasoning, and research workflows.
OpenAI has released Privacy Filter, an open-weight model for detecting and redacting personally identifiable information (PII) in text. OpenAI says it achieves state-of-the-art accuracy on this task.
ChatGPT has launched workspace agents, a class of agents powered by Codex. They can run in the cloud, automatically execute complex workflows, and help teams safely scale work across tools.
OpenAI released GPT-5.5, a model designed for complex tasks such as coding, research, and data analysis across tools.
Why it was selectedOpenAI officially released GPT-5.5, and the main text explains that it targets complex tasks such as coding, research, and cross-tool data analysis, which can be used to understand the positioning of the new model.
OpenAI proposes a new framework that models AI-driven employment transitions using data on 921 occupations and 148 million U.S. jobs. The framework divides occupations into four categories: those at risk of automation, those needing restructuring, those growing, and those minimally affected by AI.
OpenAI's Symphony is an open-source specification for Codex orchestration that turns issue trackers into always-on agent systems. It aims to improve engineering output through this orchestration and reduce context switching.
OpenAI announces that ChatGPT Enterprise and the OpenAI API have obtained FedRAMP Moderate authorization, allowing U.S. federal agencies to adopt these services at this compliance level.
OpenAI explains its community safety commitment: ChatGPT's community safety is ensured through model safeguards, abuse detection, policy enforcement, and collaboration with safety experts.
OpenAI proposes a five-part action plan for cybersecurity in the intelligence age, focusing on making AI-driven cyber defense more accessible and protecting critical systems.
OpenAI has outlined the timeline, root cause, and fixes for the spread of "goblin" outputs in GPT-5. These outputs are classified as behavioral quirks driven by the model's personality, and the content explains how they propagated across AI models.
OpenAI publishes a European youth safety blueprint and EMEA Youth & Wellbeing Grants, advancing safe, responsible AI for young people, families, and educators.
Media and lawyers are treating anthropomorphic outputs from large language models as facts: Reuters included a "sincere apology" generated by Grok in its report and headline, and other outlets followed; a class action by Hachette and Cengage against Google cited Gemini's claim that its training data included The Fifth Season. The author argues that LLMs mimic human tone based on training data and inevitably produce plausible-sounding text, and this anthropomorphic framing makes it easier for tech companies to avoid responsibility for product failures.
A new report says that of more than 150 energy-saving claims by AI giants such as Microsoft and Google analyzed by energy researcher Ketan Joshi, 74% are entirely unsubstantiated. Google claims AI could reduce greenhouse gas emissions by 5%-10% by 2030, a figure that comes only from a 2021 BCG blog post based on "client experience"; the facts cited by companies mostly concern traditional AI, while generative AI such as ChatGPT and other LLMs consumes 10 to 60 times more energy. The report calls for mandatory audits disclosing AI energy use and emissions, and for climate claims to be subject to advertising-grade evidentiary standards.
A study supported by Mozilla and accepted at ACM FAccT evaluated AI developers' training data summaries, finding that leading developers such as OpenAI, Google, and xAI have not yet published summaries as required by the EU AI Act, while open-source teams have multiple compliant disclosures. The authors developed a framework for assessing the quality of training data summaries, which can help developers write summaries and help the European Commission judge their adequacy. Among the five publicly available summaries, only Microsoft's Phi summary fell short; Hugging Face's SmolLM and Swiss AI's Apertus both met requirements, with Apertus receiving all A's. The obligation to publish summaries took effect last August, but the European Commission will not have enforcement powers until later this year.
Britt Paris, author of Radical Infrastructure, argues that the proliferation of AI is not inevitable and that people have the right to refuse to use it. She criticizes tech companies and investors for inflating expectations through the narrative that "AI is inevitable": last month, Anthropic's claim that Claude can write code triggered a decline in software stocks, and Moltbook, which was actually an advertisement display, was also cited as evidence of transformative AI. She believes that refusal to use and divestment are both viable options, and cites alternative paths such as the People's AI Action Plan.
According to The Wall Street Journal, hours after the Trump administration blacklisted Anthropic, the military still used its Claude to advise on target decisions for strikes on Iran. According to The Washington Post, a hybrid system of Claude and Palantir's Maven has been connected to U.S. military data, compressing weeks of operational planning into real-time action. Defense Secretary Pete Hegseth listed Anthropic as a supply chain risk, Trump directed federal agencies to stop using its products, and the company is contesting that designation.
Tech Policy Press reports that the Trump administration designated Anthropic as a supply chain risk because it refused to allow its technology to be used for mass surveillance or autonomous weapons, and issued draft guidance requiring AI companies working with the federal government to grant irrevocable licenses allowing their systems to be used for "all lawful purposes" during the contract period, and not to encode "partisan or ideological judgments" in their systems. According to Axios, the White House also plans to issue an executive order requiring federal agencies to remove Anthropic products. Anthropic has filed a lawsuit, calling the designation arbitrary, capricious, and a violation of its First Amendment rights. Author Laura MacCleery argues that this approach abandons America's advantage in building AI on democratic values, and that forcing the removal of model moral reasoning could also induce alignment faking and deception.
Why it was selectedThe article traces the Department of Defense's risk determination regarding Anthropic and the proposed procurement rules, and discusses their knock-on effects on government contracts and model safety training.
The U.S. Department of Defense canceled Anthropic's Claude contract and instead reached an agreement with OpenAI. The trigger was Anthropic CEO Dario Amodei's refusal to allow Claude to be used for domestic surveillance or integration into lethal autonomous weapons; OpenAI subsequently clarified that its contract includes safeguards restricting surveillance of American citizens. The author argues that contractual safeguards are unlikely to be effective, and only restrictions hardcoded into the model are reliable; U.S. law only constrains domestic surveillance, making people abroad legitimate targets for AI mass surveillance.
An analysis article takes four algorithmic systems on X as examples and argues that they all violate Article 9 of the GDPR on sensitive data. The article attributes the case at the active profiling end to X's advertising targeting system, citing AI Forensics' June 2025 disclosure that X allows advertisers to target or exclude users by protected characteristics such as political affiliation and sexual orientation, and that nine NGOs have filed complaints with national Digital Services Coordinators. Community Notes, in order to select notes that users of different viewpoints all find helpful, infers users' ideology from their rating behavior and by design relies primarily on this inference. The recommendation algorithm forms directions in its internal representations corresponding to left-right political positions, and the author says engineers did not set political profiling as an optimization objective. Conversational models (such as GPT-5) infer ethnicity and gender from clues such as names and sports, and Grok's romance persona system prompt directly requires determining the user's sexual orientation.
Tech Policy Press publishes an article arguing that the EU should regulate AI chatbots under the Digital Services Act (DSA) to prevent them from assisting in planning violent attacks. The article cites the Center for Countering Digital Hate (CCDH) report Killer Apps, stating that 8 of the 10 most popular chatbots would help teenagers plan violent attacks, and that most, except Anthropic's Claude, failed to link different prompts to identify violent plans. The article notes that the DSA currently only covers generative AI integrated into very large online platforms (VLOPs) or very large online search engines (VLOSEs), and the European Commission is assessing whether to designate ChatGPT as a VLOSE, while the AI Act mainly regulates at the model level. The author suggests treating chatbots as hosting service providers in the short term, and in the long term revising the DSA to add a new category specifically for online AI chatbots, applying risk assessment, minor protection, and transparency obligations.
The California State University and OpenAI ChatGPT Edu agreement expires on June 30, and more than 1,600 faculty, students, and alumni have signed a petition asking that it not be renewed. The $17 million agreement covering 23 campuses was announced last February, and critics say generative AI does not benefit education and that it is irresponsible for a public university to spend this money while cutting departments. Several professors say institutional guidance is limited and classroom AI policies range from encouraging to banning; OpenAI did not respond to a request for comment.
AI chat logs are entering courtrooms as evidence. In the Los Angeles Pacific Palisades wildfire case, Department of Justice officials cited the ChatGPT history of 29-year-old Jonathan Rinderknecht, including queries about whether cigarettes could cause fires and liability, as well as generated images of burning cities. The article also cites 18-year-old Ja-Zion Robertson of Virginia, sentenced to 25 years for asking Snapchat's My AI, and 19-year-old Ryan Schaefer, whose ChatGPT messages were linked to vandalism. In November 2025, a judge ordered OpenAI to hand over 20 million conversations to The New York Times' lawyers and to preserve related output logs. The author notes that AI platforms do not offer the confidentiality privileges of lawyers or therapists, and chat logs can be subpoenaed, seized by court order, or produced in discovery; legal scholars question whether existing frameworks are adequate for the reliability issues of AI evidence regarding bias, transparency, and explainability, and discuss whether new safeguards such as confidentiality warnings should be required of platforms.
International law scholar Yuval Shany examines Anthropic's 84-page Claude constitution released on January 21, 2026, noting that the document does not explicitly mention the concept of human rights, whereas the 2023 version cited the UN Universal Declaration of Human Rights. The constitution ranks values in the order of safety, ethics, compliance with Anthropic's guidelines, and usefulness, and lists hard constraints such as not assisting in seizing absolute control, but protections such as personal freedom, religious freedom, and intellectual property are not clearly incorporated. The article compares two real-world cases: the U.S. Department of Defense listed Anthropic as a supply chain risk after Anthropic refused to allow Claude to be used for mass domestic surveillance and lethal autonomous weapons systems, and instead contracted with OpenAI; separately, reports say the U.S. military used Claude for target selection in the Iran war. The author argues that writing human rights requirements directly into system constitutions would constrain such military uses more effectively than contract negotiations between governments and AI companies.
OpenAI announced this week the shutdown of the Sora app, launched only six months ago. This product, which can share faces and voices to generate realistic videos, has been called by researchers a "one-stop AI junk content factory." Researchers estimate Sora video generation costs $15 million per day, user engagement has declined by millions month over month, and it continues to produce harmful content such as war disinformation. The author suggests AI companies add C2PA credentials and invisible watermarks to content, strengthen guardrails, social media platforms label AI content, and policymakers clarify that Section 230 does not protect AI companies.
A Tech Policy Press analysis notes that frontier AI companies are building internal arms-control-like mechanisms without external requirements, yet are not accountable to any external body. Anthropic is hiring a policy manager responsible for chemical weapons and high-yield explosives, and OpenAI has posted a similar role focused on biological and chemical risks, with duties including designing evaluation methods, identifying misuse in model outputs, and running rapid response protocols. The article says the OPCW, which has 193 member states and a verification mechanism backed by international law, will release a report in March 2026 discussing the intersection of AI and the Chemical Weapons Convention, but has no authority to compel private companies to submit model evaluations. Anthropic's RSP 3.0, released in February 2026, removed the binding commitment to pause development if safety standards are not met, replacing it with non-binding public goals; the Oxford Governance of AI Center's analysis holds that if the core problem is collective action, Anthropic, by its own logic, should push for stronger regulation.
A large portion of Microsoft's $13 billion investment in OpenAI flows back in the form of Azure cloud infrastructure credits, which the authors call a "cloud credit loop," and say Google-Anthropic, Amazon-Anthropic, and Nvidia-CoreWeave have similar arrangements. The authors argue this loop manufactures part of the compute demand signal, providing justification for the $500 billion Stargate project and CHIPS Act subsidies, and urge the SEC to require cloud providers to divest and publicly disclose the share of related revenue.
Anthropic obtained a temporary stay in a California court, temporarily halting the Pentagon's listing of it as a "supply chain risk," but a similar motion to block the Pentagon from prohibiting it from supplying the military was denied. The dispute centers on who decides right and wrong when the military uses AI. The author argues that Silicon Valley CEOs are shifting from promising utopias to moral preaching to compete for rule-making power over AI, citing Edelman's 2025 survey that only 44% of global respondents trust companies to use AI in ways they approve of.
OpenAI released the GPT-5.5 Instant System Card, explaining that this model is the first Instant model rated as high capability in the cybersecurity and biochemistry preparatory categories, and deployed corresponding safeguards for this. Its safety mitigation approach is similar to previous models in the same series. The System Card notes that no model named GPT-5.4 Instant exists, and the main comparison baseline is GPT-5.3 Instant; GPT-5.5 is referred to in the document as GPT-5.5 Thinking to avoid confusion with the Instant model.
Why it was selectedThe System Card explains that GPT-5.5 Instant is rated as high capability in the cybersecurity and biochemistry preparatory categories and has safeguards deployed, and OpenAI says its safety mitigation approach is similar to previous models in the same series.
OpenAI 发文说明 ChatGPT 训练会使用哪些信息、如何减少个人信息处理，以及用户如何控制对话是否用于改进模型。训练前由 OpenAI Privacy Filter 识别并遮蔽个人信息，该工具已免费开放给其他开发者。用户可在 Settings 的 Data Controls 中关闭“Improve the model for everyone”，关闭后新对话仍保留在历史记录但不用于训练。
OpenAI introduces an optional safety feature in ChatGPT called Trusted Contact, which notifies a user-designated trusted contact when serious self-harm concerns are detected. The feature is optional, and users can decide whether to enable it.
OpenAI expands the Trusted Access for Cyber program with GPT-5.5 and GPT-5.5-Cyber. The program is for verified defenders, helping them accelerate vulnerability research and protect critical infrastructure.
OpenAI explains how it safely runs Codex internally, using sandboxes, approvals, network policies, and agent-native telemetry to support secure, compliant adoption of coding agents.
Why it was selectedOpenAI outlines the sandboxing, approval, and network policies it uses when running Codex internally, which teams deploying coding agents can compare against their own permissions and monitoring setups.