Newsletter 25

Meet Our New Campus Representatives 🔊

Our campus representatives will work to raise awareness about AI safety, organize events, and connect with interested students at their universities.

If you are at one of these universities and are interested in AI safety, you can reach out directly to our representatives.

Hatice Kübra Çetin - Boğaziçi University

Hatice, a sophomore studying Management Information Systems at Boğaziçi University, serves as the AI Safety Turkey campus representative on her campus. She is particularly interested in the social and economic impacts of artificial intelligence and the field of AI governance.

Zehra Köse - Koç University

Zehra Köse, a double major in Industrial Engineering and International Relations at Koç University, serves as the campus representative for AI Safety Turkey. Her interests include current world politics and countries’ artificial intelligence policies.

Andrés Casas - Yıldız Technical University

Originally from Colombia, now studying Computer Engineering at YTU. With a deep passion for helping others, he actively bridges technology and community through his work with AI Safety Türkiye.

ANNOUNCEMENTS 🔊

🌟MAIA AI Safety Fundamentals Fellowship: Summer 2026🌟

Online fellowship by MIT AI Alignment covering why AI safety matters, technical safety, AI policy, misalignment, and careers in the field.

📅 Deadline: May 22, 2026

🌟PRISM AI Safety Research Fellowship: 2026🌟

4-month part-time research fellowship producing work in interpretability, governance, evaluation, and alignment, with outputs aimed at conference submissions or software releases.

📅 Deadline: May 25, 2026

🌟 The Secure Program Synthesis Fellowship🌟

Part-time research fellowship by Apart Research and Atlas Computing on formal methods, AI systems, and security, including specification, validation, and adversarial robustness.

📅 Deadline: May 26, 2026

EleutherAI Summer of Open AI Research 2026

Online research program by EleutherAI where people with little or no research experience can learn by contributing to mentor-supervised open AI research projects.

📅 Deadline: May 15, 2026

Dovetail Research Fellowship 2026

10-week mathematical AI safety research fellowship focused on agent foundations and world models. Fellows work on research projects with weekly meetings, check-ins, and an end-of-term presentation.

📅 Deadline: May 17, 2026

Effective Thesis Fellowship: Summer 2026

Online fellowship pairing students and early-career researchers with high-impact partner organizations to develop thesis projects into real-world research.

📅 Deadline: June 3, 2026

Vision Weekend United Kingdom 2026

3-day Foresight Institute conference for researchers and builders, including an AI safety track led by Apollo Research on emerging AI paradigms.

Deadline: June 5, 2026

The Secure Program Synthesis Hackathon

Join a 3-days hackathon to prototype solutions that will help with AI verification. Top teams will be invited to apply to the SPS Fellowship that follows.

📅 Deadline: May 22, 2026

IAPP AI Governance Global Europe 2026

Are you working on AI regulation and compliance? This summit in Dublin brings professionals across Europe for sharing hands-on best practices of AI Act implementation and other topics in AI governance.

TOP PICKS 📑 🎧

Do Policymakers Finally Realize What Automated AI Research Mean?

AI is automating most of coding, and one sophisticated research area that it will also automate vastly is AI research itself. If current advances in AI are too fast, it will only get faster. Jack Clark explains this in detail in his most recent blog post, a reality that experts have been warning policymakers for years.

Teaching Claude Why

Why would an AI model avoid blackmail if blackmail helped it achieve its assigned goal? Anthropic’s post uses this example to explain agentic misalignment and why alignment training may need to teach models the principles behind safe behavior, not only examples of what to do.

NEWS 🗞️

OpenAI releases GPT5.5

• OpenAI announced GPT5.5 on April 23 and updated the post on April 24 to say GPT5.5 and GPT5.5 Pro were available in the API.
• The company describes GPT5.5 as a model for long-horizon work across coding, online research, data analysis, document and spreadsheet creation, software operation, and tool use.
• OpenAI says the biggest gains are in agentic coding, computer use, knowledge work, and early scientific research, while matching GPT5.4’s per-token latency and using fewer tokens on Codex tasks.
• The release also included internal and external red-teaming, targeted testing for advanced cybersecurity and biology capabilities, and feedback from nearly 200 trusted early-access partners.

Anthropic strikes SpaceX data center deal and unveils Claude “dreaming” feature

• Reuters reported on May 6 that Anthropic reached a deal to use computing resources from SpaceX’s Colossus 1 facility in Memphis, Tennessee.
• According to Reuters, the facility houses more than 220,000 Nvidia processors and is expected to give Anthropic 300 megawatts of new capacity within a month.
• Anthropic also unveiled a Claude feature called “dreaming,” intended to help its AI systems review work between sessions, identify patterns, and update files that store user preferences and context.
• The deal is important because it connects two major AI trends at once: the race for large-scale compute and the push toward more persistent, capable coding agents such as Claude Code.

CAISI signs frontier AI security-testing agreements with Google DeepMind, Microsoft, and xAI

• NIST announced on May 5 that the Center for AI Standards and Innovation signed new agreements with Google DeepMind, Microsoft, and xAI.
• The agreements allow CAISI to conduct pre-deployment evaluations and targeted research to better assess frontier AI capabilities and advance AI security.
• NIST says the agreements enable government evaluation before public release, post-deployment assessment, and testing in classified environments.
• The announcement is a major AI safety development because CAISI says it has already completed more than 40 such evaluations and that developers frequently provide models with reduced or removed safeguards for national-security risk assessment.

EU co-legislators reach AI Act simplification deal and ban “nudifier” apps

• European Parliament and Council negotiators reached a provisional deal on May 7 to amend parts of the EU AI Act as part of the digital omnibus package.
• The agreement postpones high-risk AI obligations to December 2, 2027 for high-risk use cases, and to August 2, 2028 for AI systems used as safety components under EU sectoral legislation.
• It also sets a revised application date of December 2, 2026 for watermarking obligations on AI-generated content.
• The deal bans AI systems that create child sexual abuse material or non-consensual intimate or sexually explicit images, video, or audio; it still needs formal adoption by Parliament and Council before entering into law.

Anthropic launches Project Glasswing after identifying Claude Mythos cyber risks

• Anthropic announced Project Glasswing after observing cyber capabilities in Claude Mythos Preview, a general-purpose unreleased frontier model.
• The company says Mythos Preview shows that AI models can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.
• Anthropic says Mythos Preview has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser.
• Anthropic says it does not plan to make Claude Mythos Preview generally available, making this one of the clearest recent examples of a powerful model being restricted for safety reasons while being used under controlled access for defense.

OpenAI introduces GPT‑Rosalind for life sciences research

• OpenAI introduced GPT‑Rosalind on April 16 as a purpose-built frontier reasoning model for biology, drug discovery, and translational medicine.
• The model is optimized for scientific workflows involving chemistry, protein engineering, genomics, literature review, experimental planning, and data analysis.
• GPT‑Rosalind is available as a research preview in ChatGPT, Codex, and the API for qualified customers through OpenAI’s trusted access program, alongside a Life Sciences research plugin that connects to more than 50 scientific tools and data sources.
• OpenAI says access is limited to qualified U.S. Enterprise customers at launch, with controls around eligibility, access management, organizational governance, misuse prevention, and safeguards against biological misuse.