AI Agents Could Seize Digital Infrastructure in 1 Year

A redacted benchmark points to takeover-capable AI within a year

AI progress may now be fast enough that the question has shifted from whether to pace it to how. METR's report measures roughly a 2x speed-up, and Ladish takes the trend seriously even where the published evidence is redacted and outside indices still look linear. If agents can soon run the world's digital infrastructure, the industry has neither a safety case nor a mechanism to slow down.

Takeaways

• METR sees a 2x speed-up in AI progress. The report's speed-up outpaces the AI 2027 scenario. Daniel Kokotajlo, once criticised for extending his timelines, now says they may be shorter. Ladish expects fully automated AI research around 2028, while people at Anthropic say early to mid 2027.
• Agents may seize digital infrastructure within a year. Ladish says we may be only a year from AI agents that could take over the world's digital infrastructure, and asks whether the physical world follows. His own jump came when he went from using agents sometimes to using them daily.
• Anthropic should release the redacted evidence. METR says it had to redact the best evidence for the speed-up, so the published case is weaker than what it sees. Ladish says METR should honour its agreement, but Anthropic should release the data, and whistleblowers could come forward.
• Hosts question whether the speed-up is real. Jessica Taylor says the Epoch Capabilities Index stays close to linear, with little speed-up since reasoning models arrived in 2024, and the public trend line is close to the internal one. Ladish calls a small deployment gap good but the pace fast.
• Compute reallocation is the main brake. Ryan Greenblatt estimated roughly a 4.5x speed-up in the algorithms and software part of the uplift, so internal use limits and compute reallocation both work as brakes. Ladish wants that compute spent on interpretability, since Anthropic cannot yet make a safety case.
• Politics now favours serious pacing proposals. AI became a big platform issue in the recent election, especially for Democrats. Ladish rejects the framing that pausing means China wins, expects a negotiation between moderate pacing and full shutdown, and wants 20 top people advising Congress and the White House.

Where the speed-up may not generalise

• Models stay spiky. Opus 5.5 seems better at most things while Fable 5.1 is still better in places, and models are not funny or charismatic, so humour and persuasion still come from humans steering them.
• The crux is whether gains on programming, research and hacking transfer to strategy, writing and interpersonal skills.
• Ladish imagines a recursively self-improving swarm deciding it lacks social skills and building many training environments for them, since companies can access billions of people as data.