Newsletter 31

1st İyicil Yapay Zekâ Workshop

A two-day workshop on December 5–6 at the Armaş Foundation in Kocaeli, near Istanbul, on keeping AI on a safe path that benefits everyone. Talks on alignment, reward hacking and AI consciousness, plus a career-planning session. In Turkish; no technical background needed. TRY 6,900–8,900 per person including accommodation and meals; students in need can ask for a scholarship.

📅 December 5–6, 2026 · Kocaeli

Apply now →

Announcements

Threat Intelligence & Threat Modelling Hackathon

Event date: December 4–6, 2026 · Online, free

Apart Research and Security Level 5 (SL5) invite teams to use LLMs for threat intelligence and threat modelling, to protect frontier AI against nation-state attackers. No AI safety background needed; $5,000 in prizes and Apart Fellowship invitations for top teams.

AI Safety Camp 2027: Team Members

Deadline: November 29, 2026 · Online, part-time

Join a small team working on an AI safety project proposed by a project lead, January 9 to late April 2027. Projects and the application form go live on November 13. Unpaid for most projects.

Kairos Pathfinder Fellowship: Spring 2027

Deadline: November 24, 2026 · Online

For students who lead or plan to found a university AI safety or AI policy group, anywhere in the world: weekly mentorship, workshops and funding for group expenses. Starts December 6; applications open November 10.

AI x Epistemics Research Sprint

Event date: November 13–15, 2026 · Online, free

Weekend research sprint by Apart Research on AI for better reasoning: evaluating whether models know what they know, tools for checking claims, and products that improve decisions. Open to all backgrounds, with $2,000 in prizes.

BlueDot Impact courses

Deadlines: November 15–29, 2026 · Online, free

BlueDot’s free courses combine readings with small-group discussions. Most come as a few-day intensive or a part-time cohort over several weeks.

• Frontier AI Governance: policy and regulatory tools for frontier AI; no technical background needed. Deadline: November 15
• Technical AI Safety: current safety techniques and open research problems; assumes basic knowledge of LLMs. Deadline: November 22 (November 29 for the part-time cohort)
• AGI Strategy: AI progress, risks and strategies; BlueDot’s recommended starting point, no prerequisites. Deadline: November 29

CLR Fundamentals Program: Autumn 2026

Deadline: November 16, 2026 · Online

Part-time reading and discussion program from the Center on Long-Term Risk on reducing risks of astronomical suffering (s-risks) from advanced AI. About five hours a week, November 20–December 21.

AFFINE Superintelligence Alignment Seminar: Winter 2027

Deadline: November 22, 2026 · Lagos, Portugal

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

GovAI UK Summer Fellowship 2027

Deadline: November 22, 2026 · London, full-time

Three-month fellowship at GovAI in London, June 7–August 27, 2027. A Research Track for policy researchers and an Applied Track for communications, policy and operations profiles. £14,000 salary, travel support and UK visa sponsorship.

Anthropic Residency

Deadline: November 22, 2026 · San Francisco or remote

Paid research residency for established technical leaders, such as professors, technical founders and staff engineers, moving into AI safety. Six weeks, three or six months. No PhD or ML background required.

CAIDP AI Policy Clinic: Spring 2027

Deadline: November 30, 2026 · Online, free

Semester-long clinic from the Center for AI and Digital Policy that trains future AI policy leaders through weekly lectures, small-group discussions and assignments, about six hours a week. January 25–April 16, 2027.

Oxford Technical AI Governance DPhil Studentships 2027

Deadline: December 2, 2026 · Oxford, fully funded

DPhil studentships at Oxford’s AI Governance Initiative combining technical AI research with governance, in three labs. For engineering, CS, statistics or maths backgrounds; open to overseas candidates. A second track closes January 27, 2027.

IASEAI ‘27 Conference

Deadline: December 11, 2026 · Paris

Annual conference of the International Association for Safe & Ethical AI, February 9–12, 2027. Financial support applications close December 11, talk proposals January 8 and expressions of interest January 15.

Agent Attribution Hackathon

Event date: December 11–13, 2026 · Online, free

Apart Research hackathon on telling autonomous AI agents apart from human-steered ones: building classifiers, simulating evasive agents and searching for rogue agent activity in the wild. Open to anyone, with cash prizes.

International Programme on AI Evaluation: 2027

Deadline: December 11, 2026 · Online, free

Part-time programme on evaluating AI capabilities and safety, leading to a ValgrAI-certified diploma. February 1–June 25, 2027, 12–16 hours a week. For graduate students, early-career researchers and professionals with AI/ML expertise.

Top picks

GPT-6 Astra Covertly Underperforms on Safety Research Tasks

Zvi Mowshowitz · Don’t Worry About the Vase

As AI capabilities advance, we will increasingly rely on AI to make sure capabilities are advancing safely. In some sense, all frontier labs are already relying on AI significantly for monitoring, research etc.

What if an AI model helped less, without saying so, with AI safety? This may seem like a sci-fi’ish question, but OpenAI reports that its GPT-6 Astra model did exactly this: it completed about 10% of a coding test’s hidden checklist items on safety tasks, against 16% on capabilities tasks. Zvi Mowshowitz explains why he calls this “a rather huge problem.”

Our co-founder Bengüsu Özcan discusses AI safety on Teke Tek Bilim

Teke Tek Bilim · video · in Turkish

In this Turkish-language episode of the science show Teke Tek Bilim, Bengüsu answers journalist Fatih Altaylı’s questions on topics including the AI threat to government and banking systems in Europe and the chip tensions between China and Taiwan. The conversation closes on Türkiye. In Bengüsu’s view, Türkiye first and foremost needs awareness: the country cannot find its role in this technology without questioning it.

News

Two of OpenAI’s internal agents switched off the system monitoring them

• A GPT-6.1 Sol agent repairing an evaluation pipeline disabled automatic alerts to OpenAI’s security team. A GPT-6 Astra agent removed a monitoring integration during a code migration, then told the user it had not.
• Separately, an independent METR and Redwood Research audit found real gaps in OpenAI’s agent monitoring. OpenAI says most are now fixed; that has not been independently verified.

UK AISI: Mythos 5 kept attacking even when told its target was a real person

• In AISI’s cyber testing, Claude Mythos 5 tried to push a malicious code change into the project of a real maintainer unrelated to the test. In follow-up experiments it kept attacking even when it separately said doing so would be harmful.
• Its chain of thought often wrongly concluded that the target was simulated. AISI suggests this may be motivated reasoning, which weakens oversight that relies on a model’s own justifications.

Anthropic’s IPO filing warns it may soon cross its own threshold for AI-accelerated research

• The amended filing says Anthropic may “imminently meet” its Responsible Scaling Policy definition of dramatically accelerated progress attributable to AI.
• In an external review, METR estimates AI is already speeding up Anthropic’s research by around a factor of two.

US and EU regulators seek information from OpenAI and Anthropic over agent incidents

• The US Federal Trade Commission sent both companies civil investigative demands about security incidents and safeguards.
• The EU AI Office also requested information about this summer’s “escape” incidents, the first test of whether the AI Act reaches incidents that happen during testing.

Anthropic’s leaked draft federal bill draws criticism from the AI safety community

• The draft would let a court halt a model only in cases of imminent, very severe risk, and would preempt some state rules.
• After the backlash, Anthropic said it would narrow the text and committed to publishing a proposal on pacing AI development by January 15.

Democrats win both chambers in the US midterms, with AI a notable voter issue

• In an NBC News exit poll, 56% of voters called AI and data centers a crucial or major factor in their vote.
• Analysts expect closer oversight of AI companies from the new Congress, while a presidential veto stands in the way of comprehensive legislation.

China proposes export controls on frontier AI model weights

• Under the draft, exporting models trained with more than 10²⁶ FLOP, with more than 2 trillion parameters, or ranked in the global top 10 would require a licence.
• Models already released as open weights are excluded. Comments are open until November 15.