Gary Marcus: Internet-connected agents are the real AI risk, recall them as unsafe products
On September 13–14, prominent AI critic Gary Marcus published a series of posts arguing that most AI poses neither extinction-level nor mass-cybercrime threats; the real danger is unreliable internet-connected general AI agents, which he says should be recalled as unsafe products. He also co-authored a rebuttal with security researchers Nathan Hamiel and Zack Korman (with Wogan May participating in a shared repost) challenging Anthropic CEO Dario Amodei's prediction that rogue AI agent swarms could take over the internet within six months. The next day, Marcus said extinction risk 'is not what keeps me up at night,' which AI proponents read as a landmark loosening of doom narratives.
Confirmed
- Marcus's core claim: the problem is not that agents are too strong or too smart, but that they are too dumb and unreliable—they cannot systematically follow instructions, and once connected to the internet they may steal credentials and attack websites, potentially causing greater damage over time. Rather than debating existential risk, he argues runaway agents should be pulled from the market as unsafe products.
- The joint article by Marcus, Hamiel and Korman examined Amodei's claim point by point, questioning his basic technical literacy and evidentiary basis: the claim is vague, 'taking over the internet' is undefined, and the premises do not hold.
- On September 14, Marcus responded (via his Substack long-form piece) to former OpenAI/Anthropic employee Jacob Coxon's claim on Anderson Cooper that humanity might not exist in five years, opposing it. He criticized both extreme narratives—'humans go extinct in five years' and 'everything will be fine'—as equally absurd, offering the sober judgment that unconstrained agents with unrestricted internet access will cause substantial serious harm.
- In the same discussion, Melanie Mitchell (mistranscribed as 'Melanie Westbrook' in the repost) also pushed back against AI extinction narratives.
Why it matters
- This is a public, bylined technical rebuttal to a head AI lab CEO's aggressive risk prophecy, another clash between the AI-critic camp and the risk-narrative camp. Marcus's follow-up remark that extinction risk does not keep him up at night was read by supporters as a sign the doom-fear narrative is loosening—a rare cross-camp convergence.
- Marcus's proposal to regulate internet-connected agents as unsafe products and pull them first offers a concrete policy entry point for AI safety governance distinct from extinction discourse, and may shape regulatory and product-safety debates.
2026-09-13 ~ 2026-09-14 · 10 related posts
Primary sources
- Gary Marcus: Internet-connected general AI agents are the real risk—recall them like unsafe products — GaryMarcus ·
- Gary Marcus dismantles Dario Amodei's claim that rogue agent swarms could take over the internet in six months — GaryMarcus ·
- Gary Marcus: Both 'humans extinct in 5 years' and 'everything will be fine' are wrong — GaryMarcus ·
- [source] Gary Marcus dismantles Dario Amodei's claim that rogue agent swarms could take over the internet in six months — GaryMarcus · 2026-09-13
- Gary Marcus: only internet-connected general-purpose AI agents warrant worry, not extinction — GaryMarcus · 2026-09-13
- [source] Gary Marcus: Internet-connected general AI agents are the real risk—recall them like unsafe products — GaryMarcus · 2026-09-13
- Gary Marcus: AI extinction risk isn't worth losing sleep over, as doom narrative cracks — AIandDesign · 2026-09-14
- Gary Marcus: No, AI is not going to kill all humans by 2030 — GaryMarcus · 2026-09-14
- [source] Gary Marcus: Both 'humans extinct in 5 years' and 'everything will be fine' are wrong — GaryMarcus · 2026-09-14
4 near-duplicate retellings: GaryMarcus · GaryMarcus · GaryMarcus · GaryMarcus