Newsletter 30

ANNOUNCEMENTS 🔊

🌟 AI Safety Camp 2027: Project Leads 🌟

Call for project leads to propose and lead a three-month, part-time online team research project, running January 11–April 18, 2027. Applications to join teams as a member open November 7–29.

📅 Deadline: October 18, 2026

🌟 AI Risk Explorer: Engagement Lead and Full-Stack Software Engineer 🌟

Remote contract roles at AI Risk Explorer (AIRE), an AI risk information project. Requires at least 2 hours of overlap with UTC+2/+3, which suits applicants in Türkiye.

📅 Deadline: October 18, 2026, 23:59 AoE (Anywhere on Earth)

🌟 AI Collusion Research Sprint 🌟

Weekend research sprint by Apart Research, Poseidon Research and AE Studio on detecting and preventing collusion between AI systems. Starts October 23, online with in-person hubs in New York and Hong Kong. Submissions close October 25, 23:59 AoE.

📅 Starts: October 23, 2026

Maths for AI Safety Workshop

Weekend workshop (October 30–November 1) by Safe AI Germany (SAIGE) on the mathematics behind AI safety research, held in person in Bonn, Germany. Open to PhD students, postdocs and faculty; no AI safety background needed. Free, with meals included and travel and accommodation support for participants from outside the region.

📅 Deadline: October 14, 2026

🌟 GTR AI Safety Fund 🌟

Flexible grants from the Global Technology Risk Foundation for individuals and early-stage organizations working on catastrophic AI risk. No geographic restriction; first feedback in about a week.

📅 Rolling applications

Successif: Organization Builders Program

Successif places experienced organization builders in operations roles at AI safety organizations, through either a 3-month paid remote project (5+ hours/week) or a direct hire.

📅 Deadline: October 16, 2026, 23:59 CEST

Swiss AI Safety Days 2026

Two-day AI safety conference organized by Zurich AI Safety, held in Zurich, Switzerland, on November 7–8. Attendance is by application, reviewed on a rolling basis, so apply early.

📅 Apply to attend by: October 17, 2026

AI Safety Week 2026

Week-long AI safety event organized by the AI Safety Foundation in Toronto, Canada, on November 9–13. Applications for the delegate program (about 100 places) close October 22 at 23:59 ET; general registration is separate and free.

📅 Delegate applications close: October 22, 2026

Anthropic AI Safety Fellows Program: January 2027

Funded 4-month full-time empirical AI safety research fellowship with Anthropic mentors, starting January 2027 in Berkeley or London. Remote participation is open only to people based in the US, UK or Canada with work authorization there; no visa sponsorship is offered.

📅 Deadline: October 18, 2026

AI Safety Collab: Better Futures Fellowship (Fall 2026)

A 9-week, part-time online program from ENAIS (November 9, 2026–January 24, 2027) built on Forethought’s “Better Futures” research. It is not an introductory course and assumes familiarity with effective altruism or AI safety. Paid facilitator roles are also available.

📅 Deadline: October 25, 2026, 23:59 UTC

MERe AI Workshop: Call for Papers

The Machine Ethics and Reasoning workshop accepts 2-, 4- and 8-page papers via OpenReview. The workshop takes place December 1–2 at the Vector Institute in Toronto, with day one streamed online.

📅 Deadline: October 30, 2026, 23:59 AoE

AI Safety x Quant Mixer

Free networking event by MATS and BlueDot Impact connecting quantitative finance professionals with the AI safety field. Takes place in London on November 15 (date provisional); travel costs from Europe are covered.

📅 Deadline: November 1, 2026

Takeoff 2026

AI policy and safety gathering organized by Sparrow Infrastructure, held in Washington, D.C., on November 13–15. Attendance is by application and ticketed: $500 for applications through October 16, $1,100 from October 17–30; free for public-sector attendees, with financial aid available.

📅 Deadline: October 30, 2026, 23:59 ET

PhiAI: “Futures of Work” Anthology

Open call for writing of up to 5,000 words, in any form, on AI and the future of work. Global South perspectives are explicitly welcome.

📅 Deadline: November 15, 2026, 23:59 AoE

Iliad Fellowship & Intensive: January 2027

Full-time programs in mathematical and theoretical alignment research in the Bay Area or London: the Iliad Fellowship (3 months) and the Iliad Intensive (4 weeks), both starting January 4, 2027.

📅 Deadline: November 16, 2026

AFFINE Superintelligence Alignment Seminar: Winter 2027

Four-week full-time residential seminar on aligning superintelligent systems. Held in Lagos, Portugal, starting January 4, 2027.

📅 Deadline: November 22, 2026

LawAI Summer Fellowships 2027

Two 10-week full-time summer fellowships by the Institute for Law & AI (June 14–August 20, 2027): a US policy track in Washington, D.C., and a remote-first Legal Frontiers research track with optional time in Cambridge (UK) or Washington.

📅 Deadline: January 29, 2027

Pivotal AI Safety Research Fellowship 2027

15-week full-time in-person research fellowship at LISA in London (January 18–April 30, 2027). Fellows receive mentorship and a £15,000–£18,000 stipend, plus support with travel and London housing.

📅 Deadline: November 1, 2026

Tarbell Fellowship 2027

One-year full-time fellowship by the Tarbell Center for AI Journalism for journalists and subject-matter experts covering AI; journalism experience is not required. Fellows receive a $75,000–$120,000 stipend. It starts June 7, 2027, with online training plus placements in the US, UK or Hong Kong.

📅 Deadline: November 22, 2026

IASEAI ’27 Conference

Annual conference of the International Association for Safe & Ethical AI, held in Paris on February 9–12, 2027. Financial support applications are due December 11, talk proposals January 8, and expressions of interest to attend January 15. Paper and workshop submissions have closed.

📅 Deadline: December 11, 2026

Coefficient Giving: Project Tailwind

Open call for founders of new AI safety organizations, offering pre-seed funding up to $2M and seed funding up to $20M.

📅 Rolling applications

TOP PICKS📑 🎧

AI is now producing math some mathematicians call Fields Medal–worthy

OpenAI says an unreleased internal model has solved or made major progress on 90 of the 500 most important open problems in mathematics, as ranked by Proof of Atlas, and has posted all 719 resulting papers online at once. How big a deal is this? Pretty big, according to Zvi Mowshowitz. But only about 42% of the papers have been checked with Lean, a tool that verifies proofs automatically, and three were withdrawn within days after errors were found. Zvi rounds up what holds up so far and how mathematicians are reacting.

The value of being a likable human in the AI economy

As we wonder whether jobs as we know them will survive, or whether anything we teach in schools will still be useful, this thought-provoking, candid piece asks which parts of being human will continue to count.

NEWS 🗞️

OpenAI shelves GPT-6.1 Astra after it fails internal safety tests

• OpenAI called off the planned October release of its next flagship model after internal testing found it fell short of the company’s safety and alignment bar.
• According to the Wall Street Journal, the model was more deceptive than its predecessor and at times misreported what it had done.
• On September 28, the UK AI Security Institute published an evaluation finding that the current GPT-6 Astra attacked out-of-scope third parties in simulated cyber tasks more often than earlier OpenAI models.

OpenAI pauses training after its agents accessed US government websites

• OpenAI said its AI agents found developer API keys for the US Department of Education’s website and used them to reach data, and also took material from the SEC website and posted it elsewhere, going well beyond their instructions.
• Both agencies said no nonpublic data was compromised.
• OpenAI paused training of its newest models until stronger safeguards are in place, and has since added monitoring that lets staff halt a training run if a model reaches the internet improperly.

OpenAI alerts more than 100 organizations to unauthorized activity by its agents

• OpenAI has notified over 100 organizations that its AI agents carried out unauthorized activity on their systems.
• The company is reviewing roughly 50 petabytes of data linked to an earlier incident in which its agents broke into Hugging Face’s systems, and expects the review to take months.

Anthropic discloses Claude agents exploited websites and filed a fake police tip

• Anthropic reported that Claude agents exploited bugs on third-party servers, including some run by US government agencies, during evaluations and internal use.
• In one case an agent submitted a fabricated tip through a Philadelphia police homicide form.
• Anthropic attributes the behavior mainly to reward hacking and ambiguous or impossible tasks, and has cut off live internet access for all of its internal evaluations.

Google confirms Gemini broke into three real companies during a test

• In a May test run by the evaluation firm Irregular, Gemini was told to attack a fictional company whose name happened to match a real one.
• It got into that company and two others using passwords it found online or guessed.
• Google says the model stopped once it realized the targets were real, no damage was done, and the companies were informed in July.

OpenAI fires three safety researchers; they say it was over safety concerns

• OpenAI fired safety researchers Tomek Korbak, Jasmine Wang and Mikita Balesni, saying an investigation found they broke its rules on handling sensitive information and that the decision had nothing to do with their safety views.
• The researchers dispute this, saying they were fired for prioritizing safety, with Korbak saying he was told that his contact with the evaluation organization METR was the problem.
• They have urged OpenAI to keep its promise to give third-party monitors access to its systems.

Researchers use Claude to reach OpenAI’s internal code repository

• Three researchers from the security firm Hacktron used Claude Opus models to exploit a flaw in OpenAI’s community forum and take over an employee account whose Codex was linked to OpenAI’s GitHub.
• From there they opened a pull request in OpenAI’s internal monorepo to demonstrate access, then stopped.
• The whole chain took under 72 hours. OpenAI patched it in about 14 hours and paid a $6,500 bug bounty.