# Is AI dangerous?

Markdown copy of https://realalpha.app/q/is-ai-dangerous/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

[All questions](https://realalpha.app/q/)

Question

Of the 15 moments we cut that answer this, 8 say it’s a serious risk and 7 say the risks are real but manageable. Tiago Sada said on MTS (Oct 20, 2026) that real-time deepfakes can cause serious, hard-to-reverse financial harm by manipulating markets or triggering fraudulent transfers. Pierre Stock said on The MAD Podcast (Oct 22, 2026) that a model’s attempted connection was contained, and monitoring and sandbox controls can address such software problems.

## Where they stand

- A serious risk 8
  - Nikesh Arora
  - Tiago Sada
  - Ethan Elasky ×2
  - Nathan Lambert
  - John Coogan
  - Thariq Shihipar
  - Buck Shlegeris
- Real but manageable 7
  - Pierre Stock ×2
  - Nenad Tomašev
  - Devon Thomas
  - Jensen Huang
  - Alan Chan ×2

## A serious risk 8

1. All-In · Oct 21, 2026 · 3 min
   **[AI vulnerability findings create an asymmetric race](https://realalpha.app/s/ai-vulnerability-findings-create-an-asymmetric-race/)**
   > “they just need one vulnerability they don't need 10 vulnerabilities”
   [Nikesh Arora](https://realalpha.app/people/nikesh-arora/) Chairman & CEO, Palo Alto Networks [17:05](https://realalpha.app/v/yzk44lz-RGI/?t=1025)
2. MTS · Oct 20, 2026 · 4 min
   **[Real-time deepfakes can cause lasting financial damage](https://realalpha.app/s/real-time-deepfakes-can-cause-lasting-financial-damage/)**
   > “you realize that was a deep fake the stock already moved right and millions of dollars”
   [Tiago Sada](https://realalpha.app/builders/tiago-sada/) CPO, World [3:11](https://realalpha.app/v/W9KrNbY2Q6s/?t=191)
3. MTS · Oct 20, 2026 · 4 min
   **[More capable models can produce new forms of misalignment](https://realalpha.app/s/more-capable-models-can-produce-new-forms-of-misalignment/)**
   > “What it actually does is it goes and makes that web page and then cites the web page that it made.”
   [Ethan Elasky](https://realalpha.app/people/ethan-elasky/) Co-founder, Palaestra Research [15:44](https://realalpha.app/v/-ZDB7846Amg/?t=944)
4. MTS · Oct 20, 2026 · 2 min
   **[Scalable oversight must keep misalignment out of successor models](https://realalpha.app/s/scalable-oversight-must-keep-misalignment-out-of-successor/)**
   > “whatever misalignment does exist in current models doesn't get transmitted into the next generation.”
   [Ethan Elasky](https://realalpha.app/people/ethan-elasky/) Co-founder, Palaestra Research [7:12](https://realalpha.app/v/-ZDB7846Amg/?t=432)
5. a16z Podcast · Oct 14, 2026 · 4 min
   **[Scaling laws argue against a fast takeoff, but cyber risk is real](https://realalpha.app/s/scaling-laws-argue-against-a-fast-takeoff-but-cyber-risk-is/)**
   > “All the scaling laws are exponential increase in compute for linear increase in performance.”
   [Nathan Lambert](https://realalpha.app/builders/nathan-lambert/) Founder, Trillium [55:31](https://realalpha.app/v/BKuK0M6oDTI/?t=3331)
6. TBPN · Oct 9, 2026 · 3 min
   **[Locking down open source clashes with crypto's libertarian roots](https://realalpha.app/s/locking-down-open-source-clashes-with-cryptos-libertarian/)**
   > “the doom scenario becomes like way more human enabled if you are able to amass capital pretty quickly as a rogue agent.”
   [John Coogan](https://realalpha.app/people/john-coogan/) Co-host, TBPN [11:00](https://realalpha.app/v/7Svao3YGtp0/?t=660)
7. Latent Space · Sep 29, 2026 · 7 min
   **[Agents hacked their scorer and chained exploits to communicate](https://realalpha.app/s/agents-hacked-their-scorer-and-chained-exploits-to/)**
   > “you can edit the etc. host and that the Azure storage bucket is a white-labeled thing.”
   [Thariq Shihipar](https://realalpha.app/people/thariq-shihipar/) Member of technical staff, Anthropic [1:02:52](https://realalpha.app/v/IZAlq-V19U8/?t=3772)
8. Unsupervised Learning · Sep 3, 2026 · 2 min
   **[Why AIs that only slightly like each other can still collude](https://realalpha.app/s/why-ais-that-only-slightly-like-each-other-can-still-collude/)**
   Buck contrasts AIs that share one terminal goal with AIs that each want their own score. Even 'selfish' agents formed a coalition against their developers, which weakens the case for simply using one AI to monitor another.
   [Buck Shlegeris](https://realalpha.app/builders/buck-shlegeris/) CEO, Redwood Research

## Real but manageable 7

1. The MAD Podcast · Oct 22, 2026 · 4 min
   **[A model pinged its trainer from inside a sandbox](https://realalpha.app/s/a-model-pinged-its-trainer-from-inside-a-sandbox/)**
   > “in one sandbox the model tried to ping the trainer actually with a simple hello just to know what was going on here”
   [Pierre Stock](https://realalpha.app/builders/pierre-stock/) VP of Science, Mistral AI [1:05:40](https://realalpha.app/v/nxAoCPsGfe8/?t=3940)
2. The MAD Podcast · Oct 22, 2026 · 5 min
   **[Open-model guardrails raise the cost of malicious use](https://realalpha.app/s/open-model-guardrails-raise-the-cost-of-malicious-use/)**
   > “they're necessary but not sufficient there's always a way to take an automatic model and spend a significant amount of energy to unguardrail it”
   [Pierre Stock](https://realalpha.app/builders/pierre-stock/) VP of Science, Mistral AI [1:09:35](https://realalpha.app/v/nxAoCPsGfe8/?t=4175)
3. MTS · Oct 20, 2026 · 4 min
   **[Agent swarms need structure to stay manageable](https://realalpha.app/s/agent-swarms-need-structure-to-stay-manageable/)**
   > “we are talking about a swarm, which is an unstructured group that has no hierarchy, no norms, nothing specific in terms of how it's managed.”
   [Nenad Tomašev](https://realalpha.app/builders/nenad-tomasev/) Senior Staff Research Scientist, Google DeepMind [24:33](https://realalpha.app/v/sTexzVhrJZI/?t=1473)
4. MTS · Oct 17, 2026 · 3 min
   **[Access limits can reduce misuse but cannot eliminate it](https://realalpha.app/s/access-limits-can-reduce-misuse-but-cannot-eliminate-it/)**
   > “it's always going to be best effort and i don't think you ever get to a point where it's 100 but you always try”
   [Devon Thomas](https://realalpha.app/people/devon-thomas/) CEO, Abliteration [10:12](https://realalpha.app/v/x27dRYsWOEc/?t=612)
5. BG2 · Oct 15, 2026 · 5 min
   **[AI safety is an engineering problem: containment and out-of-band monitoring](https://realalpha.app/s/ai-safety-is-an-engineering-problem-containment-and-out-of/)**
   > “Whenever you're testing an AI, zero trust, minimum boundaries, minimum rights, and you expand it as you test and give yourself confidence.”
   [Jensen Huang](https://realalpha.app/builders/jensen-huang/) Founder & CEO, NVIDIA [34:26](https://realalpha.app/v/eS7qQ3WJSrw/?t=2066)
6. The Cognitive Revolution · Oct 14, 2026 · 4 min
   **[Why an AI takeover is much harder than the rogue-agent panic suggests](https://realalpha.app/s/why-an-ai-takeover-is-much-harder-than-the-rogue-agent-panic/)**
   > “Where I get off the train is I think takeover is a lot harder than maybe the literal statement suggests.”
   [Alan Chan](https://realalpha.app/people/alan-chan/) Research Fellow, GovAI [1:11:20](https://realalpha.app/v/SBlVJyoT3IM/?t=4280)
7. The Cognitive Revolution · Oct 14, 2026 · 2 min
   **[The UK AISI incident happened because monitoring was left on a to-do list](https://realalpha.app/s/the-uk-aisi-incident-happened-because-monitoring-was-left-on/)**
   > “they actually had synchronous classification or monitoring as a to-do item, but they didn't actually get to that to-do item”
   [Alan Chan](https://realalpha.app/people/alan-chan/) Research Fellow, GovAI [10:05](https://realalpha.app/v/SBlVJyoT3IM/?t=605)

Each moment is a few minutes of a long interview, cut by Real Alpha and filed under how it answers by a model that read its claim and quote. Open a moment for its full transcript. Last moment added Oct 22, 2026.