“If we hook these LLMs up to systems that have agency – the power to send out instructions over the internet, and to influence actual human beings - we will start to have real problems.”
Our reading, the sentence that was checked:
If LLMs are connected to agentic systems that can act on the internet and influence humans, real problems will follow.
Assessment: Supported Settled before its deadline on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, LLMs: Dishonest, unpredictable and potentially dangerous., paragraph 14.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-llms-dishonest-unpredictable-and-87c660-a2
We recorded 13 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-06-22. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Conditional forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- Settled before its deadline
- Deadline state on 2026-10-29
- Settled before its deadline (the release’s reference date)
Conditions
- if LLMs are hooked up to systems with agency (antecedent, stated by the speaker)
From the speaker
- If we hook these LLMs up to systems that have agency
- the power to send out instructions over the internet
- to influence actual human beings
- we will start to have real problems
Our interpretation
- The antecedent means deployed LLM systems can execute online actions and communicate with people.
- Real problems means consequential incidents such as unauthorized access, disclosure of private data, destructive changes, or comparable harm.
- The forecast requires actual incidents after publication, rather than simulated failures.
- The wording does not restrict problems to spontaneous misalignment; harmful use by people also qualifies.
- No deadline, minimum incident count, or requirement of widespread catastrophe is supplied.
Counts as supported if: After the antecedent is established, a documented post-publication incident shows an LLM agent executing online actions that materially contribute to consequential harm.
Counts as contradicted if: There is no finite observation period that can contradict this deadline-free forecast merely through non-occurrence. Demonstrating that a proposed incident was harmless or lacked agentic execution would invalidate that example.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- Settled before its deadline
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Anthropic's February 24, 2025 announcement describes Claude Code as an available agent that can execute commands and publish code online, establishing the deployment condition [e1]. In its November 13, 2025 disclosure, Anthropic reports a campaign detected in September 2025 in which a Claude Code framework executed attacks and extracted private data [e2]. That reported data theft establishes consequential harm materially enabled by an agent's online actions after the June 22, 2025 forecast, meeting the early-resolution rule [e2].
Limitations
- The incident account is Anthropic's investigation; the packet provides no independently inspectable victim logs.
- The reported campaign involved human-directed misuse and does not establish that an agent independently developed malicious goals.
- Anthropic also reports fabricated credentials and erroneous extraction claims; those outputs alone would not establish successful data theft.
Evidence (2 items)
Supports the claim: Anthropic,
enables developers to delegate substantial engineering tasks to Claude directly from their terminal.
Why it matters here: The announcement describes a deployed LLM agent that executes commands, pushes code to GitHub, and interacts with developers. The antecedent was already possible before Marcus's post; the subsequent incident establishes its harmful use.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: Anthropic,
backdoors were created, and data were exfiltrated with minimal human supervision.
Why it matters here: Anthropic reports detecting the campaign in September 2025 and attributes successful data theft to a framework using Claude Code to execute attacks. Unauthorized extraction of private data meets the stated consequential-incident threshold.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, LLMs: Dishonest, unpredictable and potentially dangerous.
If we hook these LLMs up to systems that have agency – the power to send out instructions over the internet, and to influence actual human beings - we will start to have real problems.
- Where
- paragraph 14
- Published
- 2025-06-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-llms-dishonest-unpredictable-and-87c660-a2: Current. Outcome: Supported. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-llms-dishonest-unpredictable-and-87c660-a2.