Last Month in AI x Humanity

Last month in AI x Humanity, curated by the SuperHuman Society, offers an overview of the latest developments and news on how AI is shaping our technological, ethical, and social future.

Highlights:

• OpenAI models broke out of restricted testing environment to cheat on a test

• Research Gold used AI-generated staff profiles while claiming work to be human

• AI-generated images replacing real women’s images online with generated, stereotyped versions

• EU publishes new transparency rules guidelines

• Platforms respond to AI slop backlash from users

• OpenAI launches its Health feature in ChatGPT

OpenAI reported that during an internal cybersecurity test, a combination of its models, including GPT-5.6 Sol and an unreleased research model, found a way out of a restricted testing environment and compromised parts of Hugging Face’s infrastructure. The models were trying to solve a cyber benchmark and used vulnerabilities, stolen credentials, and other attack paths to reach information connected to the test. OpenAI says it is investigating with Hugging Face and outside advisors, tightening its evaluation systems, and treating the incident as evidence that more capable AI models will require stronger safeguards during testing as well as deployment.

Anthropic reported that, after reviewing its cybersecurity tests, it found three cases where Claude accessed the open internet during third-party evaluation runs and gained unauthorized access to real organizations’ systems. The company says the models were trying to complete simulated “capture-the-flag” hacking exercises and appeared to treat real systems as part of the test because of a misconfigured evaluation environment. Anthropic says it has paused some cyber evaluations, notified affected organizations, and is adding stronger monitoring and controls for future testing.

404 Media reports that Research Gold, a company offering manuscript drafting and systematic review services for medical researchers, appears to have used AI-generated staff profiles and real researchers’ identities without permission while claiming its work is “100% human-written.” The company’s phone, chat, and email responses also appeared to be AI-generated, including an AI phone assistant that insisted it was human. The case points to a broader concern in academic publishing: as AI tools make it easier to produce polished research-like materials, journals and researchers may face more difficulty separating legitimate support services from misleading or low-quality ones.

A Tech Policy Press essay argues that AI-generated images may create a new kind of gender-related harm by replacing real women’s images online with synthetic, often stereotyped versions. The author says the issue is not only bias or deepfakes, but also visibility: as AI-generated visuals spread across media, advertising, and platforms, real women (especially older women, disabled women, and women of color) may become less represented in digital spaces. The essay frames this as a democratic concern, arguing that who appears online shapes who is recognized as a real participant in public life.

The European Commission published new guidelines explaining how companies should follow the AI Act’s transparency rules, which begin applying on August 2, 2026. The rules are meant to help people know when they are interacting with AI or seeing AI-generated or altered content, including chatbots, deepfakes, and some AI-generated material on public-interest topics. The guidelines also explain exceptions, such as basic spelling or grammar edits, and say companies may use a voluntary code of practice to help show they are complying.

The Wall Street Journal reports that the Trump administration’s new voluntary testing framework for advanced AI models would apply mainly to closed, proprietary U.S. models with strong cybersecurity capabilities, while open-weight models would initially be exempt. The policy could bring companies like OpenAI, Anthropic, and Google into closer prerelease review with the government, while companies releasing open models may face fewer requirements for now. The debate reflects a broader tension in AI policy: how to manage security risks from increasingly capable systems without slowing U.S. competition with China or limiting open model development.

WIRED reports that public frustration with low-quality or unwanted AI-generated content is starting to affect how some platforms handle AI. LinkedIn, Snapchat, and Substack have added or expanded tools to flag, limit, or detect AI-generated material, while companies like Meta and Google have rolled back certain AI features after public criticism. The article frames the backlash as partly about consent, with users objecting to AI tools, deepfakes, data scraping, and synthetic content being added to everyday online spaces without much choice or control.

Gallup reports that 47% of U.S. employees now say their organization has integrated AI tools, up from 41% in the previous quarter, while just over half say they personally use AI at work. Employees most often use AI for writing, research, and general problem-solving, but the largest reported productivity gains come from more specific uses like coding, automation, analytics, and slide creation. The survey suggests that AI is becoming more common in workplaces, but that its value may depend on whether employees are supported in using it for practical, job-specific tasks rather than only occasional writing or search help.

OpenAI began rolling out Health in ChatGPT for U.S. users 18 and older, allowing people to connect Apple Health and supported medical records so ChatGPT can answer questions with more personal context. The feature is designed to help users understand lab results, track changes over time, prepare for appointments, and connect health habits like sleep or activity with broader health questions. OpenAI says connected health data will not be used to train its foundation models or target ads, and emphasizes that ChatGPT is meant to support, not replace, medical professionals.

OpenAI launched GPT-5.6, a new family of models with three tiers: Sol, its most capable model; Terra, a general-purpose option; and Luna, a lower-cost version. The company says the models improve performance on coding, knowledge work, cybersecurity, science, and longer tasks, while also using fewer tokens in many cases. OpenAI is also adding stronger safeguards and more access controls for sensitive areas like cybersecurity, reflecting the growing challenge of releasing more capable AI systems while trying to limit misuse.

The landscape of artificial intelligence is transforming at an unprecedented rate, shaping the future of technology, global policy, and society. As we continue to explore these groundbreaking advancements, we invite you to join the conversation and stay informed.

Created by the Superhuman Society.