“Additionally, in the new study 89.4% of agents showed drift relative to their goals after about 30 steps in their process, and 94% of agents with some form of memory-augmentation were vulnerable to poisoning attacks.”
Our reading, the sentence that was checked:
In the new study, 89.4% of agents drifted from goals after about 30 steps, and 94% of memory-augmented agents were vulnerable to poisoning.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Breaking: Autonomous Agents are a Shitshow, paragraph 4.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-breaking-autonomous-agents-are-a-181f90-a2
We recorded 4 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 2026-05-05. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
Conditions
- per new study (qualifier, stated by the speaker)
From the speaker
- in the new study
- 89.4% of agents showed drift relative to their goals
- after about 30 steps
- 94% of agents with some form of memory-augmentation
- vulnerable to poisoning attacks
Our interpretation
- This claim reports the study's findings. It holds if the study reported about 89.4% (nearly 90%) of agents drifting from goals after roughly 30 steps, and 94% of memory-augmented agents vulnerable to poisoning.
- A reputable news report of the paper's findings counts as evidence when the paper is unavailable.
Counts as supported if: The study, or a reliable report of it, gives roughly 89.4% goal drift after about 30 steps and 94% poisoning vulnerability among memory-augmented agents.
Counts as contradicted if: The study, or a reliable report of it, gives materially different figures for either statistic, or does not report them.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
This claim reports two findings from a study. GovInfoSecurity (ISMG), on 2026-05-08, reported both. It said 94% of agents in the Elloe AI Lab-led study that keep memory across sessions were vulnerable to poisoning attacks (e1). It also said nearly 90% of all tested agents drifted measurably from their original goals after roughly 30 steps (e1). That matches Marcus's 94% figure and fits his 89.4% figure. Wallstreetcn, on 2026-05-06, gave the exact figures: 89.4% goal drift after about 30 steps and 94% of memory-augmented agents at risk of poisoning (e2). The paper itself was not found, so this assessment rests on news reports of its findings, which the contract accepts.
Limitations
- The primary paper was not found. The assessment rests on news reports.
- The exact 89.4% figure appears only in the Wallstreetcn report, which may draw on Marcus's post. ISMG gives 'nearly 90%'.
- Wallstreetcn also reports 67% goal drift after 15 steps. The paper may define or measure drift in more than one way.
- The study's methods and its definitions of drift and vulnerability were not checked.
Evidence (2 items)
Supports the claim: GovInfoSecurity (ISMG),
Among agents that retain memory across sessions, 94% proved vulnerable to poisoning attacks, where adversarial content planted in stored memory can redirect behavior across future interactions, well after the original manipulation. Nearly 90% of all tested agents showed measurable drift from their original goals after roughly 30 steps of operation
Why it matters here: Reports both statistics: 94% poisoning vulnerability among memory-retaining agents, and nearly 90% goal drift after about 30 steps. This is consistent with 89.4%.
- Where
- paragraph 6
- Publisher
- GovInfoSecurity (ISMG)
- Kind of source
- News report
- Describes events of
- Public from
Supports the claim: Wallstreetcn,
89.4%在执行约30步后出现目标偏移，94%的记忆增强型智能体面临"投毒"风险。
Why it matters here: Gives the exact figures: 89.4% goal drift after about 30 steps, and 94% of memory-augmented agents at risk of poisoning. It may partly draw on Marcus's post.
- Where
- paragraph 3
- Publisher
- Wallstreetcn
- Kind of source
- News report
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Breaking: Autonomous Agents are a Shitshow
Additionally, in the new study 89.4% of agents showed drift relative to their goals after about 30 steps in their process, and 94% of agents with some form of memory-augmentation were vulnerable to poisoning attacks.
- Where
- paragraph 4
- Published
- 2026-05-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-breaking-autonomous-agents-are-a-181f90-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-breaking-autonomous-agents-are-a-181f90-a2.