“Jailbreaks aren’t new, but even after years of them, the tech industry has nothing like a robust response.”
Our reading, the sentence that was checked:
Despite years of jailbreaks, the tech industry has no robust defense against them.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, When it comes to security, LLMs are like Swiss cheese — and that’s going to cause huge problems, paragraph 17.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-when-it-comes-to-security-llms-are-19b54d-a1
We recorded 7 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 2024-10-17. 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)
From the speaker
- Jailbreaks aren’t new
- even after years of them
- the tech industry has nothing like a robust response
Our interpretation
- The ordinary reading is that available defenses did not reliably withstand varied, realistic jailbreak attacks.
- Robust does not mean perfect resistance to every conceivable attack.
- The statement does not deny the existence of mitigations, successful defenses against particular attacks, or ongoing research.
- For the ongoing state without a specified period, use the latest completed calendar quarter before the post, July through September 2024. Earlier work supplies historical context.
Counts as supported if: Contemporary evaluations show substantial failures of credible defenses under realistic attack variations, supporting the absence of an established broadly reliable response.
Counts as contradicted if: Contemporary evaluations establish an available, practical defense that reliably resists varied adaptive jailbreaks while preserving useful operation. Success against one fixed attack alone is insufficient.
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
On the ordinary reading of robust as reliable against varied realistic attacks, the contemporary evaluation supports Marcus's assessment. Defenses that performed well on automated single-turn tests remained substantially vulnerable to multi-turn human attacks. Effective attack-specific mitigations existed, so the evidence would not support a stronger claim that industry had done nothing or that every defense was useless.
Limitations
- The result depends on the stated reading of robust.
- The principal study evaluates four defenses, not every commercial deployment.
- The reporting quarter is an evaluator choice.
- This assesses the 2024 state, not defenses introduced later.
Evidence (2 items)
Supports the claim: arXiv; Scale AI and UC Berkeley researchers,
fail to generalize
Why it matters here: The study evaluates four defenses and reports substantial vulnerability to human multi-turn attacks despite strong results against automated single-turn attacks. This supports the ordinary reading that established defenses lacked broad robustness.
- Publisher
- arXiv; Scale AI and UC Berkeley researchers
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Anthropic,
implemented mitigations
Why it matters here: Anthropic reports a mitigation reducing one attack's success from 61% to 2%, while also describing weaknesses and possible evasions. This is material evidence against reading Marcus as saying no effective response existed at all; it predates the target quarter and does not establish broad robustness.
- 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, When it comes to security, LLMs are like Swiss cheese — and that’s going to cause huge problems
Jailbreaks aren’t new, but even after years of them, the tech industry has nothing like a robust response.
- Where
- paragraph 17
- Published
- 2024-10-17
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-when-it-comes-to-security-llms-are-19b54d-a1: 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-when-it-comes-to-security-llms-are-19b54d-a1.