As of: 2026-10-18 07:35 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/v/imf-ai-and-escalating-cyber-risks-panel-sellitto/

# Artificial Intelligence and Escalating Cyber Risks

IMF, 16 October 2026, YouTube. 3,055 views as of 18 October 2026.
Kind: Official.
Watch: https://www.youtube.com/watch?v=FyPANjJEfN8

## Why it is here

The IMF's own recording (60 min) of the Oct 16 Annual Meetings panel on AI and cyber risk in Bangkok, with Anthropic's Michael Sellitto, FINMA's Marlene Amstad, the BIS's Andrea Maechler and the Central Bank of Brazil's Paulo Picchetti. It is the primary source for Sellitto's remark that Claude led 26% of building Anthropic's next model in August and that the share "could be 80%, 90%" next year, which spread on X as "80–90% of frontier model training".

## Description (written by Gemini from the video)

**Summary**  
An official panel discussion hosted by the International Monetary Fund (IMF) during the 2026 Annual Meetings previews and examines the escalating cybersecurity and systemic financial stability risks posed by advanced artificial intelligence. Moderated by Bo Li (Deputy Managing Director, IMF), the panel features Michael Sellitto (Head of APAC Policy, Anthropic), Marlene Amstad (Chair, Swiss Financial Market Supervisory Authority), Paulo Picchetti (Deputy Governor, Central Bank of Brazil), and Andrea Maechler (Deputy General Manager, Bank for International Settlements). The participants debate offensive versus defensive AI dynamics, model self-improvement velocity, third-party infrastructure vulnerabilities, and the urgent need for international regulatory coordination.

**What is shown**  
- [00:45] Bo Li introduces the session and framing questions around macro-financial stability, systemic shock propagation, and regulatory responses to AI-accelerated cyber threats.
- [04:40] Michael Sellitto discusses frontier model capability scaling, internal automation at Anthropic, the withheld release of Claude Mythos, and results from Project Glasswing.
- [14:15] Marlene Amstad outlines three systemic risk transmission channels (core application disruption, shared infrastructure/vendor outages like MOVEit, and confidence/liquidity runs driven by perception) along with SupTech supervisory initiatives.
- [27:10] Paulo Picchetti discusses lessons from attacks targeting external service providers connected to Brazil’s Pix payment network and how cyber intrusions have largely replaced physical bank robberies.
- [32:20] Andrea Maechler analyzes interconnectedness, the collapsing window between vulnerability discovery and exploitation, and the challenge of decision latency for central bank crisis response.
- [40:48] Second-round panel discussion examining containment, restoration, trade-offs between efficiency and resilience (including potential "no-fly zones" for AI in core functions), and the BIS "digital twin" simulation program.
- [58:50] Lightning closing round where panelists identify priority actions for the coming year, emphasizing supply-chain transparency and international cooperation.

**Claims & numbers**  
- **Internal coding automation:** Michael Sellitto states that at Anthropic, the proportion of internal code written by Claude went from 0% early last year to ~50% in August, and reached 80%–90% by the end of last year [06:41].
- **AI self-training acceleration:** Sellitto claims that the share of the next-generation model training process led by Claude rose from 0% at the start of the year to 3% in April and 26% in August, projecting it could reach 80%–90% next year [07:35].
- **Project Glasswing findings:** Anthropic and its cybersecurity partners discovered ~130,000 vulnerabilities during the 90-day Glasswing initiative, of which ~30,000 were rated critical or severe, with the vast majority remaining unpatched months later [09:52].
- **Attacker efficiency:** Breaches that previously required a team of human specialists weeks or months can now be executed by a single person in hours using coordinated agent workflows [11:00].
- **Swiss supervisory survey:** Marlene Amstad notes FINMA surveyed over 200 banks, nearly 200 insurers, 10,000 funds, and 150 platforms regarding AI integration [15:18].
- **Decline in physical bank robberies:** Paulo Picchetti states physical bank robberies in Brazil have fallen to almost zero as criminal actors shifted toward cheaper, lower-risk cyber fraud and credential extortion [29:43].

**Notable quotes**  
- [07:38] *Michael Sellitto:* "At the beginning of the year it was 0%, in April it was 3%, in August 26% of the process of building the next-generation model is being led by Claude."
- [33:43] *Andrea Maechler:* "The interesting thing with AI cyber risk is that it's neither an AI risk nor is it a cyber risk really... frontier AI turns a simple cyber incident really into a potential financial stability issue."
- [34:58] *Andrea Maechler:* "The time between the discovery and the exploitation is collapsing super fast... and we need to really be able to build up and close that gap."

**Assessment**  
This is an official IMF panel discussion and policy briefing, recorded live on stage in front of an audience. No software or live cyber tools are directly demonstrated on screen; the session consists of senior policymakers and industry representatives presenting empirical observations, internal operational metrics, and institutional regulatory strategies.

_Described by gemini-3.8-flash on 2026-10-18 from the video's audio and frames._

## Related

- 2026-10-16: [Sellitto at the IMF: Claude could lead 80–90% of model building next year](https://postcutoff.com/e/2026-10-16-sellitto-imf-claude-80-90-percent-model-training/)

What else happened that month, day by day: https://postcutoff.com/timeline/2026/10/