MIT Technology Review · Jan 11, 2025
Jared Kaplan argues in MIT Technology Review that agents need safer, broader capabilities
Markdown · JSON · Original source
On January 11, 2025, MIT Technology Review published an interview with cofounder and chief scientist Jared Kaplan about how agents may change. Kaplan argues that agents could become more useful by working across different tools and environments, learning users’ context, and improving coding assistance. He expects computer use to help models handle more complex tasks and recognise mistakes, while context about a person’s role and constraints could help do what users expect. These are Kaplan’s expectations, not reported outcomes. Kaplan also warns that broader use brings safety challenges. He names , in which malicious prompts can reach a model through sources such as websites, as a particular concern for computer use. He says Anthropic is actively working on the risk and that more capable models could be misused. The interview closes with both Kaplan’s excitement about AI’s usefulness, including its use within Anthropic, and his expectation of significant challenges. The article notes that computer use remains cumbersome and error-prone; it does not establish whether the anticipated improvements or safeguards will materialise.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
MIT Technology Review · Article
Source details
Original source date: Jan 11, 2025 (2025-01-11T11:00:00Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-16T01:13:25.928Z. First observed: 2026-11-01T00:03:40.371Z. Retrieved: 2026-11-01T00:05:21.944Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on November 1, 2026; originally published January 11, 2025. The interview reports Kaplan’s expectations; it provides no results demonstrating whether the anticipated improvements or safeguards will occur. The article describes computer use as cumbersome and error-prone.
Open question
How will address as computer use expands, and how will it assess whether its safeguards work?
Source-stated
Kaplan expects agents to improve at using tools and tackling more complex tasks, and to better recognise mistakes and when to seek user feedback.
computer use will probably improve quickly in terms of how well models can do different tasks and more complex tasks. And also to realize when they’ve made mistakes
Source-stated
Kaplan says agents need context about users’ situations and constraints to be useful and safe.
Claude needs to learn enough about your particular situation and the constraints that you operate under to be useful.
Source-stated
Kaplan identifies prompt injection as a major concern for broader agent use, particularly computer use.
Prompt injection is probably one of the No.1 things we’re thinking about in terms of, like, broader usage of agents.
Source-stated
Kaplan says Anthropic is actively working on prompt injection risks associated with computer use.
it’s something we’re working on very actively, because if computer use is deployed at large scale
Source-stated
The article says Anthropic’s computer use feature is still cumbersome and error-prone.
Anthropic notes that the feature is still cumbersome and error-prone.
Source-stated
Kaplan says more advanced models carry greater risks, including potential misuse.
And with more advanced models, there’s just more risk. We have a robust scaling policy where, as AI systems become sufficiently capable, we feel like we need to be able to really prevent them from being misused.
Source and contributor context
MIT Technology Review · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Prompt injection
Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 116c9346-40a5-aa72-2b38-bf6fa2e3967e.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T01:09:32.574Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: 66e1a1137b3e4a8e386afb21d07cda6c10c316742000b21f761f38199f1e4eaf. Source text hash: 0cc955e8c1fc9654e550d78f901351185dce4e51f2ff433428fc2ef0b5cb02cf.
Revision and source-check history
Version 1 · 2026-11-16T01:13:25.928Z
Jared Kaplan argues in MIT Technology Review that agents need safer, broader capabilities
On January 11, 2025, MIT Technology Review published an interview with cofounder and chief scientist Jared Kaplan about how agents may change. Kaplan argues that agents could become more useful by working across different tools and environments, learning users’ context, and improving coding assistance. He expects computer use to help models handle more complex tasks and recognise mistakes, while context about a person’s role and constraints could help do what users expect. These are Kaplan’s expectations, not reported outcomes. Kaplan also warns that broader use brings safety challenges. He names , in which malicious prompts can reach a model through sources such as websites, as a particular concern for computer use. He says Anthropic is actively working on the risk and that more capable models could be misused. The interview closes with both Kaplan’s excitement about AI’s usefulness, including its use within Anthropic, and his expectation of significant challenges. The article notes that computer use remains cumbersome and error-prone; it does not establish whether the anticipated improvements or safeguards will materialise.
Original source
MIT Technology Review · Article
Source details
Original source date: Jan 11, 2025 (2025-01-11T11:00:00Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-16T01:13:25.928Z. First observed: 2026-11-01T00:03:40.371Z. Retrieved: 2026-11-01T00:05:21.944Z.
Limits
Text as retrieved on November 1, 2026; originally published January 11, 2025. The interview reports Kaplan’s expectations; it provides no results demonstrating whether the anticipated improvements or safeguards will occur. The article describes computer use as cumbersome and error-prone.
Open question
How will address as computer use expands, and how will it assess whether its safeguards work?
Source-stated
Kaplan expects agents to improve at using tools and tackling more complex tasks, and to better recognise mistakes and when to seek user feedback.
computer use will probably improve quickly in terms of how well models can do different tasks and more complex tasks. And also to realize when they’ve made mistakes
Source-stated
Kaplan says agents need context about users’ situations and constraints to be useful and safe.
Claude needs to learn enough about your particular situation and the constraints that you operate under to be useful.
Source-stated
Kaplan identifies prompt injection as a major concern for broader agent use, particularly computer use.
Prompt injection is probably one of the No.1 things we’re thinking about in terms of, like, broader usage of agents.
Source-stated
Kaplan says Anthropic is actively working on prompt injection risks associated with computer use.
it’s something we’re working on very actively, because if computer use is deployed at large scale
Source-stated
The article says Anthropic’s computer use feature is still cumbersome and error-prone.
Anthropic notes that the feature is still cumbersome and error-prone.
Source-stated
Kaplan says more advanced models carry greater risks, including potential misuse.
And with more advanced models, there’s just more risk. We have a robust scaling policy where, as AI systems become sufficiently capable, we feel like we need to be able to really prevent them from being misused.
Source and contributor context
MIT Technology Review · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Prompt injection
Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: 116c9346-40a5-aa72-2b38-bf6fa2e3967e.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T01:09:32.574Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: 66e1a1137b3e4a8e386afb21d07cda6c10c316742000b21f761f38199f1e4eaf. Source text hash: 0cc955e8c1fc9654e550d78f901351185dce4e51f2ff433428fc2ef0b5cb02cf.
uncertain · 2026-11-16T01:09:32.574Z · anthropic/claude-opus-5.5
Publication: 116c9346-40a5-aa72-2b38-bf6fa2e3967e
attribution — pass: This is a single MIT Technology Review interview with Anthropic cofounder and chief scientist Jared Kaplan. The reading attributes the forecasts to Kaplan and the cumbersome, error-prone description to the article or Anthropic.
MIT Technology Review talked to Anthropic’s cofounder and chief scientist Jared Kaplan
summary — uncertain: The reading tracks Kaplan's four themes and keeps both his optimism and his warnings. A general reader is still left without two central terms. 'Computer use' is never explained; the source describes it as an extension of Claude that operates a computer by moving a cursor, clicking buttons and typing. 'Agents' and 'Claude' (Anthropic's model) are also undefined. Smallest repair: add one clause defining computer use and agents, and one identifying Claude as Anthropic's model.
it lets you direct Claude to use a computer much as a person would, by moving a cursor, clicking buttons, and typing text.
date — pass: January 11, 2025 is the UTC calendar day of originalPublished. Forecasts are framed as Kaplan's expectations, not outcomes. The later sidebar items in the captured page (Deep Dive stories) are neither used nor relied on. The retrieval date is disclosed, and this is a publisher page, not an archive copy.
support — pass: Each claim is supported by quoted source text: tool and task improvement, recognising mistakes, the need for context, prompt injection as a top concern, active work on that risk, misuse risk in more advanced models, and the cumbersome, error-prone status of computer use. The closing question follows from the evidence without presuming an answer. The modification date does not affect the cited interview text.
Prompt injection is probably one of the No.1 things we’re thinking about in terms of, like, broader usage of agents.
media — pass: No media was selected, and no verified candidates were supplied, so nothing was misidentified.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeHidden instructions placed in a document, email or web page that try to take control of an AI system reading it.
Definition referenceAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicHidden instructions placed in a document, email or web page that try to take control of an AI system reading it.
Definition referenceAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeHidden instructions placed in a document, email or web page that try to take control of an AI system reading it.
Definition referenceAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicHidden instructions placed in a document, email or web page that try to take control of an AI system reading it.
Definition reference