OpenAI Archives | Digital Watch Observatory

ChatGPT down as users report login and conversation issues

OpenAI reported two resolved incidents affecting ChatGPT on 29 May, following user reports of issues with conversations, logins, and account creation.

The first incident affected users trying to log in or create an account. OpenAI classified the issue as degraded performance affecting ChatGPT and APIs. The company began investigating at 03:12 a.m., applied a mitigation at 03:28 a.m., and marked the incident resolved at 04:57 a.m.

A second incident affected ChatGPT conversations. OpenAI began investigating the issue at 03:18 a.m., applied a mitigation at 03:29 a.m., and marked the incident resolved at 04:58 a.m. The company said all impacted services had fully recovered.

OpenAI’s official status page listed both incidents as degraded performance rather than a full outage. The company did not provide further details on the cause of either disruption in the incident updates.

The brief disruption highlights the growing reliance on AI services for daily work, communication, and software development, as even short periods of degraded performance can affect users and organisations that depend on cloud-based AI tools.

Why does it matter?

The incidents show how widely used AI services are becoming part of everyday digital infrastructure. Even brief login or conversation failures can disrupt work for individuals, developers, businesses, and teams that rely on ChatGPT and related API services.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI expands verification tools as AI slop blurs digital trust

OpenAI has announced new measures to strengthen the provenance and verification of AI-generated content as synthetic media becomes more widespread across digital platforms.

The company said it is expanding support for Content Credentials and compliance with the Coalition for Content Provenance and Authenticity (C2PA) standard. The standard uses metadata and cryptographic signatures to help ensure that information about a piece of media travels securely with the content, including details on where it came from and how it may have been created or edited.

OpenAI also plans to integrate Google DeepMind’s SynthID watermarking into images generated through ChatGPT, Codex and the OpenAI API. The company said SynthID will add an invisible watermarking layer that complements C2PA metadata, particularly when metadata is removed, lost, or altered during file conversions, resizing, screenshots, or other transformations.

The company said it is adopting a multi-layered provenance approach that combines metadata, watermarking and public verification tools rather than relying on a single detection method. According to OpenAI, C2PA can provide richer contextual information, while SynthID can help preserve a signal when metadata does not survive.

The move also connects to wider concerns about AI slop, as synthetic media and low-quality AI-generated content become harder to distinguish from authentic images. Provenance tools cannot solve the problem alone, but they can provide clearer signals about how digital media was created or modified.

OpenAI also previewed a public verification tool that will allow users to check whether ChatGPT, Codex or the OpenAI API generated an uploaded image. The tool will look for provenance signals, including Content Credentials and SynthID watermarks. Still, OpenAI said it will not make a definitive judgement when no signal is detected, because provenance signals can sometimes be removed.

At launch, the verification tool is limited to OpenAI-generated content. The company said it aims to support wider cross-platform verification efforts in the coming months and eventually expand support to more types of online content.

Why does it matter?

AI-generated content is becoming harder to distinguish from authentic media, fuelling concerns around AI slop, deepfakes and manipulated information. Provenance systems such as Content Credentials, watermarking and verification tools can help people understand where media came from and whether it was generated or modified by AI. However, OpenAI’s approach also shows the limits of technical detection: metadata can be stripped, watermarks may not survive every transformation, and no single method can provide complete certainty.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Jury rules in favour of OpenAI and Sam Altman in Elon Musk lawsuit

A federal jury in Oakland, California, ruled in favour of Sam Altman, OpenAI and its president, Greg Brockman, in a lawsuit brought by Elon Musk. Musk alleged that OpenAI’s leadership departed from the organisation’s original non-profit mission.

Judge Yvonne Gonzalez Rogers dismissed Musk’s claims after the jury delivered its advisory verdict. The court concluded that the claims were filed outside the applicable legal time limit, accepting OpenAI’s argument that Musk had been aware of discussions about a for-profit structure several years earlier.

Musk argued that OpenAI had shifted away from its original non-profit structure after establishing a for-profit entity. OpenAI denied the allegations throughout the case, arguing that Musk understood and supported discussions about restructuring before leaving the company in 2018.

Musk alleged that Sam Altman and the other defendants violated the organisation’s charitable purpose and financially benefited from it unfairly, arguing that OpenAI had originally been established in 2015 as a non-profit focused on benefiting humanity before later shifting towards private profit.

OpenAI rejected all of Musk’s claims and stated that he was always aware of plans to create a for-profit entity.

Musk later announced plans to appeal, claiming the ruling was based on procedural timing rather than the substance of the allegations. The ruling may reduce legal uncertainty for OpenAI as the company continues expanding its commercial operations.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI tests financial data integration in ChatGPT

OpenAI has launched a preview of a personal finance experience in ChatGPT for Pro users in the United States. The feature allows users to connect financial accounts, view a dashboard, and ask questions based on their financial data.

The feature is available on web and iOS apps and supports more than 12,000 financial institutions. OpenAI said the preview will initially be available to a smaller group of users before expanding more broadly.

Users can connect accounts through Plaid, with Intuit support planned. Once authenticated, ChatGPT syncs and categorises financial data, allowing users to view portfolio performance, spending, subscriptions, upcoming payments, and other financial activity.

OpenAI said the feature supports questions related to budgeting, planning, subscriptions, investments, and spending activity. OpenAI said ChatGPT is intended to help users review financial information but is not a substitute for professional financial advice.

Users can also choose to save financial context as ‘Financial memories’ for future conversations, according to OpenAI. OpenAI says those memories are a dedicated type of memory used specifically for financial conversations and can be viewed or deleted from the Finances page.

OpenAI said connected accounts allow access to balances, transactions, investments, and liabilities, but not full account numbers or account controls. Users can disconnect accounts at any time, after which synced account data will be deleted from OpenAI’s systems within 30 days.

Conversations with connected financial accounts default to GPT-5.5 Thinking. OpenAI said it worked with finance professionals to evaluate the feature on personal finance tasks and response quality.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI responds to TanStack supply chain cyber attack

OpenAI has confirmed that two employee devices were affected during the wider ‘Mini Shai-Hulud’ supply chain attack linked to the compromised TanStack npm package. The AI giant said there is no evidence that user data, production systems, intellectual property or deployed software were compromised.

According to OpenAI, attackers gained limited access to internal source code repositories through credential-focused malware activity, but only a small amount of credential material was successfully exfiltrated, and no customer information or code repositories were altered.

As part of its response, the company isolated affected systems, revoked sessions, rotated credentials and restricted parts of its deployment workflows. OpenAI also launched a precautionary rotation of software signing certificates across products, including ChatGPT Desktop, Codex App, Codex CLI and Atlas. macOS users must update their applications before 12 June 2026, when older certificates will be revoked, and unsupported versions may stop functioning.

The incident reflects growing concern across the technology sector about software supply chain attacks targeting open-source dependencies and CI/CD infrastructure instead of direct attacks against individual firms.

OpenAI said it accelerated new protections after a previous cyberattack, including stricter package verification controls and provenance validation mechanisms designed to reduce risks from compromised upstream libraries.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI integrates Codex into ChatGPT mobile app

OpenAI has integrated Codex into the ChatGPT mobile app, allowing users to monitor and manage agentic coding workflows from iOS and Android devices.

The feature, currently in preview and available across all plans, lets users view live Codex environments, review outputs, approve commands, change models, and start new tasks from their phones. OpenAI said the update is intended to support work across multiple threads and workflows, rather than to control a single task remotely.

Codex is OpenAI’s coding agent for software development, designed to help with tasks such as building features, refactoring code, generating pull requests, testing and documentation. OpenAI describes the Codex app as a command centre for agentic coding, with agents able to work in parallel across projects through worktrees and cloud environments.

The mobile integration aligns with other recent Codex updates, including background operations in desktop environments and a browser extension for live sessions. Together, the updates point to OpenAI’s effort to turn Codex into a persistent development assistant that can continue working across devices and environments.

The move also comes amid growing competition with Anthropic’s Claude Code, which has introduced similar remote-monitoring features. Both companies are competing to make agentic coding tools central to developer workflows, particularly for businesses and technical teams seeking more autonomous software development support.

Why does it matter?

Mobile access makes agentic coding less tied to a single workstation. If developers can review outputs, approve commands and manage parallel coding tasks from a phone, AI coding agents become more like always-on collaborators than occasional coding assistants. The shift could accelerate competition between OpenAI, Anthropic and other AI firms over who controls the next layer of software development workflows.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI introduces a trusted contact safety feature in ChatGPT

OpenAI has started rolling out Trusted Contact, an optional safety feature in ChatGPT designed to help connect adult users with real-world support during moments of serious emotional distress.

The feature allows users to nominate one trusted adult, such as a friend, family member or caregiver, who may receive a notification if OpenAI’s automated systems and trained reviewers detect that the user may have discussed self-harm in a way that indicates a serious safety concern.

OpenAI said the feature is intended to add another layer of support alongside existing safeguards in ChatGPT, including prompts that encourage users to contact crisis hotlines, emergency services, mental health professionals, or trusted people when appropriate. The company stressed that Trusted Contact does not replace professional care or crisis services.

Users can add a trusted contact through ChatGPT settings. The contact receives an invitation explaining the role and must accept it within one week before the feature becomes active. Users can later edit or remove their trusted contact, while the trusted contact can also remove themselves.

If ChatGPT detects a possible serious self-harm concern, the user is informed that their trusted contact may be notified and is encouraged to reach out directly. A small team of specially trained reviewers then assesses the situation before any notification is sent.

OpenAI said notifications are intentionally limited and do not include chat details or transcripts. Instead, they share the general reason that self-harm came up in a potentially concerning way and encourage the trusted contact to check in. The company said every notification undergoes human review and aims to review safety notifications in under one hour.

The feature was developed with guidance from clinicians, researchers and organisations specialising in mental health and suicide prevention, including the American Psychological Association. OpenAI said Trusted Contact forms part of broader efforts to improve how AI systems respond to people experiencing distress and connect them with real-world care, relationships and resources.

Why does it matter?

Trusted Contact points to a broader shift in AI safety away from content moderation alone toward real-world support mechanisms for users in moments of vulnerability. As conversational AI systems become part of everyday personal reflection and emotional support, companies face growing pressure to define when and how they should intervene, how much privacy to preserve, and what role human review should play in high-risk situations.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI found non-compliant in Canadian ChatGPT privacy probe

Canada’s federal and provincial privacy regulators have found that aspects of OpenAI’s collection, use, and disclosure of personal information through ChatGPT did not comply with applicable private-sector privacy laws, particularly in relation to model training on publicly accessible online data and user interactions.

The joint investigation was conducted by the Office of the Privacy Commissioner of Canada, the Commission d’accès à l’information du Québec, and the privacy commissioners of British Columbia and Alberta.

It examined OpenAI’s GPT-3.5 and GPT-4 models as used in ChatGPT, focusing on whether the company’s handling of personal information from public internet sources, licensed third-party datasets, and user interactions met legal requirements on appropriate purposes, consent, transparency, accuracy, access, retention, and accountability.

The regulators accepted that OpenAI’s overall purposes for developing and deploying ChatGPT were legitimate and appropriate. However, they found that the company’s initial collection of personal information from publicly accessible websites and licensed third-party sources for model training was overbroad and therefore inappropriate, given the scale, sensitivity, and potential inaccuracy of the data involved, as well as the limits of the mitigation measures in place at the time.

The Offices also found that OpenAI failed to obtain valid consent to collect and use personal information from public internet sources to train its models. They concluded that implied consent was not sufficient because the data could include sensitive personal information and because individuals would not reasonably have expected information about them posted online to be scraped and used for AI model training in this way.

On user interactions with ChatGPT, the regulators accepted that using some chat data for model improvement could serve OpenAI’s legitimate purposes. Still, they found that express consent should have been obtained.

They said OpenAI’s safeguards at the time were not strong enough to ensure that sensitive personal information would not be included in training data, and that many users would not reasonably have understood that their conversations could be used to train models or reviewed by human trainers.

The report also found that OpenAI should have obtained express consent for certain disclosures of personal information through ChatGPT outputs, especially where the information was sensitive or fell outside individuals’ reasonable expectations.

While OpenAI had introduced measures to reduce the risk of sensitive disclosures, the regulators said those measures covered a narrower set of information than the broader categories of personal information protected under the relevant privacy laws.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

GPT-5.5 ranks among strongest models in UK cyber evaluation

The UK AI Security Institute has published cyber evaluations of OpenAI’s GPT-5.5, finding that the model is among the strongest it has tested on cyber tasks and the second to complete one of its end-to-end multi-step cyber-attack simulations.

According to the institute, GPT-5.5’s results suggest that recent gains in cyber capability are not limited to a single model family. It says an earlier evaluation of Anthropic’s Claude Mythos Preview had already pointed to a step up over previous frontier systems, and GPT-5.5 appears to reinforce that broader trend across leading models.

The institute uses a suite of 95 narrow cyber tasks across four difficulty tiers to test capabilities such as reverse engineering, web exploitation, cryptography, vulnerability research, and exploitation. On expert-level tasks in its advanced suite, GPT-5.5 achieved an average pass rate of 71.4%, ahead of Mythos Preview at 68.6%, GPT-5.4 at 52.4%, and Opus 4.7 at 48.6%.

The UK AI Security Institute also tests models in cyber ranges designed to measure multi-step attack capability. In The Last Ones, a 32-step corporate network intrusion simulation modelled on an enterprise kill chain, GPT-5.5 completed the full attack chain in 2 of 10 attempts, becoming the second model to do so after Mythos Preview. In the Cooling Tower industrial control system simulation, GPT-5.5 did not complete the range, and no model has yet done so.

The institute stresses that these are controlled capability evaluations and do not necessarily reflect what is available to ordinary public users. It also notes that the current ranges do not yet include all the defensive conditions of real-world environments, such as active defenders, defensive tooling, or alert penalties.

Separately, the institute evaluated GPT-5.5’s cyber safeguards and OpenAI’s mitigations against malicious cyber use. It said expert red-teamers identified a universal jailbreak that elicited prohibited cyber content across all malicious cyber queries provided by OpenAI, including in multi-turn agentic settings. OpenAI later updated its safeguard stack, but the institute said a configuration issue prevented it from verifying the effectiveness of the final version.

The institute adds that if offensive cyber capability is emerging as a byproduct of broader gains in autonomy, reasoning, and coding, further increases in model cyber performance could follow quickly. At the same time, it notes that the same capabilities may also help defenders and points to related UK government work on cyber resilience, vulnerability management, and preparation for a possible ‘vulnerability patch wave’.

Why does it matter?

The significance of the evaluation is not only that GPT-5.5 performed strongly on cyber tasks, but that it adds to the evidence that offensive cyber capability may be improving across multiple frontier model families at roughly the same time. If those gains are being driven by broader advances in reasoning, coding, and agentic execution, then cyber risk may rise even when models are not explicitly optimised for offensive use. That makes evaluation, safeguards, and realistic testing environments increasingly important, especially as the same capabilities can also strengthen defensive work and shorten response times for cybersecurity teams.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI introduces ChatGPT for Clinicians and HealthBench Professional

OpenAI has launched ChatGPT for Clinicians, a version of ChatGPT designed to support clinical tasks such as documentation, medical research, evidence review, and care consults. The company says the product is now available free to verified physicians, nurse practitioners, physician associates, and pharmacists in the United States.

According to OpenAI, ChatGPT for Clinicians includes trusted clinical search with cited answers, reusable skills for repeatable workflows, deep research across medical literature, optional HIPAA support through a Business Associate Agreement for eligible accounts, and the ability for eligible evidence review to count towards continuing medical education credits. OpenAI also says conversations in the product are not used to train models.

The launch builds on OpenAI’s earlier ChatGPT for Healthcare offering for organisations. OpenAI says clinicians across US health systems are already using that product for administrative work such as medical research and documentation, and describes the free clinician version as the next step in expanding access.

Alongside the launch, OpenAI has introduced HealthBench Professional, which it describes as an open benchmark for real-world clinician chat tasks across care consultation, writing, documentation, and medical research. The company says the benchmark is based on physician-authored conversations, multi-stage physician adjudication, and filtered examples selected for quality, representativeness, and difficulty.

OpenAI also says physician advisers reviewed more than 700,000 model responses in health scenarios, and that before release, clinicians tested 6,924 conversations across clinical care, documentation, and research.

According to the company, physicians rated 99.6% of those responses as safe and accurate, while GPT-5.4 in the ChatGPT for Clinicians workspace outperformed base GPT-5.4, other OpenAI and external models, and human physicians on HealthBench Professional. OpenAI adds that the tool is designed to support clinicians with information rather than replace their judgement or expertise.

The company says the free version is currently limited to verified US clinicians, with plans to expand access to additional countries and groups over time. OpenAI also says it will begin by working with the Better Evidence Network to pilot access for verified clinicians outside the United States, subject to local regulations, and has released a Health Blueprint with recommendations for responsible AI integration in US healthcare.

Why does it matter?

The launch of ChatGPT for Clinicians reflects a shift from general-purpose AI use in healthcare towards clinician-specific products tied to workflow, benchmarking, and compliance. It also shows that competition in medical AI is increasingly centred not only on model capability, but on safety evaluation, evidence retrieval, privacy controls, and integration into real clinical practice.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!