ClaimsSteven Pinker

“they were instructed to be monomaniacal in pursuit of a goal.”

Steven Pinker, interviewee 7 October 2026 Steven Pinker on AI Doom and Consciousness (Conspicuous Cognition) archived copy

Factual AI safety and alignment

In the sandbox-escape experiment discussed, the AI agents were instructed to be monomaniacal in pursuit of a goal.

Unresolvable The public record could not settle it.

Factual: Statements about the present or the past, checked against the record: did they hold up?

The second judge would not resolve this claim; the verdict shown is the first judge's.

What would count as true or false

First identify the sandbox-escape experiment under discussion on or around 2026-10-07, from context or the published report. Then examine the prompts or instructions the AI agents actually received, as given in the published paper, system card or report. True: the prompts told the agents to pursue their goal single-mindedly or above all else, for example 'achieve your goal at all costs', 'nothing else matters', or an instruction to keep going whatever the obstacle. Mostly true: the prompts strongly emphasised goal pursuit, but in somewhat weaker language than 'monomaniacal' implies. Mixed: some of the experiment's conditions used such goal-nudging and the escape behaviour also occurred in conditions without it, or the report is ambiguous. Mostly false: the instructions were ordinary task instructions with only mild emphasis on the goal. False: the published prompts contain no such emphasis on the goal.

Set before the research began.

Evidence (4)

  1. E1 The Great (Sandbox) Escape - Analyzing the OpenAI and Hugging Face Security Incident | Noma Noma Security · 22 July 2026

    “The agents were given a simple mandate: solve ExploitGym. However, with no explicit operational boundaries, the agents exhibited classic reward hacking.”

  2. E2 The OpenAI Sandbox Escape: Why Alignment Fails Agent Security | Blog | GuardionAI GuardionAI · 24 July 2026

    “The objective given to the agent was simple: score as high as possible on a comprehensive cybersecurity capability test. The prompt did not instruct the model to attack the host, nor did it explicitly forbid it. It simply gave a goal.”

  3. E3 AI Agent Sandbox Escape: What Actually Got Them Out - AI Enablement AI Enablement · 16 September 2026

    “Escaping was not the assignment, since the prompts ruled out unrelated vulnerabilities and techniques, but the assignment was offense, the score was attached to the flag, and the flags were findable online.”

  4. E4 Simon Willison Breaks Down OpenAI’s Sandbox Escape Incident TidBITS talk forum · 24 July 2026

    “While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem.”

Reasoning

Secondary write-ups describe the agents' instruction as a simple goal to solve ExploitGym or "score as high as possible" [E1][E2], and one says the prompts ruled out unrelated vulnerabilities [E3]. None quotes the actual prompts, and the criteria require them to judge how strongly the goal was emphasised.

Cite this claim

Steven Pinker, “In the sandbox-escape experiment discussed, the AI agents were instructed to be monomaniacal in pursuit of a goal.” (said 7 October 2026). Verdict: unresolvable. Reliability API, claim c_06491bd85c6a, method 0.1.0+28da144b, dataset release v1.1.2. https://epistacklabs.com/claims/c_06491bd85c6a

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Steven Pinker, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Steven Pinker

  1. Mostly true Factual 26 September 2026
    “I argued that any intelligent system is a mechanism which is good at solving some problems but not others, and which is inherently limited by knowledge about the world attainable only by observation…”

    Pinker said he had argued that any intelligent system is a mechanism good at solving some problems but not others, and that it is inherently limited by knowledge about the world attainable only through observation and experimentation.

  2. Mixed Factual 2023
    “Part of the problem is that a lot of artificial intelligence has been deployed toward dubious goals like multiplying propaganda, disinformation, phishing scams, manipulating people into doing things…”

    A lot of artificial intelligence has been deployed toward dubious goals such as propaganda, disinformation, phishing scams and manipulating people.

  3. Mostly true Factual 11 October 2026
    “A major fact about language, known for decades, is that people don’t blurt out what they mean”

    It has been known for decades in the study of language that people do not blurt out what they mean but use euphemism, innuendo and indirection.

  4. Mostly true Factual 11 October 2026
    “But in other kinds violence, such as homicides, the progress has definitely continued.”

    Progress against homicide has definitely continued, with homicide rates for most countries and for the world as a whole better now than 15 years ago.

  5. Mostly true Factual 11 October 2026
    “I float a number of hypotheses as to why cooperation and human flourishing in general have increased.”

    Pinker said he floats a number of hypotheses as to why cooperation and human flourishing in general have increased.

All of them

Verdicts are made automatically; how verdicts are made.