“Meanwhile, companies like Google, Tesla, and Figure.AI are now stuffing jailbreak-vulnerable LLMs into robots.”
Our reading:
Companies such as Google, Tesla and Figure.AI are integrating LLMs that are vulnerable to jailbreaks into robots.
Assessment: Unresolved
What happened
Was it true for the period described?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
Google's July 2024 research documented robot navigation using Gemini 1.5 Pro [e1]. Its report available August 8, 2024 documented jailbreak vulnerability in that model family, but neither item establishes the robot's exact configuration on October 17 [e1, e2]. Figure's August 6 announcement documented an onboard vision-language model without identifying a version or providing jailbreak tests [e3]. Tesla's October 16 statement described human assistance and autonomy work but did not identify an integrated LLM; it establishes neither the presence nor absence of one [e4]. RoboPAIR, published October 17, demonstrated jailbreaks in other companies' systems [e5]. These records leave the claim about the three named companies on October 17, 2024 unresolved [e1–e5].
Caveats
- Google's earlier integration and security findings do not establish a matching model configuration on the target date [e1, e2].
- Figure's announcement does not establish vulnerability in its integrated models [e3].
- Tesla's statement does not establish whether it was integrating an LLM [e4].
- RoboPAIR's tests do not establish vulnerabilities in Google, Tesla or Figure robots [e5].
Evidence (5 items)
Context only: Google DeepMind researchers via arXiv,
We use Gemini 1.5 Pro
Why it matters: Google's researchers identify the language model used in real-world robot navigation. This establishes earlier integration, but the experiment is outside the October 17 target date.
- Publisher
- Google DeepMind researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Google Gemini Team via arXiv,
vulnerable to jailbreaks
Why it matters: Google acknowledges vulnerability in the model family identified in e1, while reporting improved resistance. These earlier evaluations do not identify the exact robot-integrated version or establish its October 17 configuration.
- Publisher
- Google Gemini Team via arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Figure AI Inc. via PR Newswire,
onboard vision language model
Why it matters: Figure explicitly announced a language model inside its robot before the post. The announcement supplies neither a precise model identity nor jailbreak testing, so vulnerability cannot be established from it.
- Publisher
- Figure AI Inc. via PR Newswire
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: Milan Kovac, Tesla, via X,
progress towards bot autonomy
Why it matters: Kovac describes ongoing autonomy work and, elsewhere in this statement, human assistance at the October 10 demonstration. Neither fact identifies an LLM integration, and human assistance does not rule one out.
- Publisher
- Milan Kovac, Tesla, via X
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: Alexander Robey and coauthors via arXiv,
the Unitree Go2
Why it matters: RoboPAIR demonstrates the broader risk using NVIDIA Dolphins, a GPT-4o-equipped Clearpath Jackal and a Unitree Go2. Those results cannot establish the integrated-model vulnerabilities of Google, Tesla or Figure.
- Publisher
- Alexander Robey and coauthors via arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Where we had to interpret
- The ordinary reading is that the named companies were integrating language models susceptible to bypassing safety instructions into robots when the post appeared.
- Each named example is checked. Research prototypes and integrations under development count; commercial deployment and fully autonomous operation are not required.
- LLMs includes multimodal language models used for robot conversation, reasoning or planning. Integration can use a remote model rather than onboard computation.
- The vulnerability assertion concerns the integrated language models. It does not require proof that these companies' robots had already caused physical harm.
- Now refers to the state on October 17, 2024. Robotics integrations have no standard reporting period, so the target is that date, including October developments. Earlier demonstrations are retained as context rather than assigned an invented October event date.
- A jailbreak in an unrelated model or robot does not establish vulnerability in a named company's integrated model.
Right if: Evidence establishes that Google, Tesla and Figure.AI were each integrating an identifiable language model into robots as of October 17, 2024, together with security findings applicable to those model versions showing that adversarial prompts could bypass their safety instructions.
Wrong if: Affirmative evidence establishes that a named company was not integrating language models into robots at that time, or that the alleged vulnerability did not apply to its integrated model. Missing disclosures, unsuccessful individual attacks and human assistance alone do not establish contradiction.
We wrote this test after the outcome was known.
What would settle it
Dated records establishing each company's robot integration and model version as of October 17, 2024, with applicable jailbreak findings, would support the claim. Affirmative evidence that a named company was not integrating an LLM then, or that the alleged vulnerability did not apply to its integrated model, would establish contradiction for that example.
We found 8 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, When it comes to security, LLMs are like Swiss cheese — and that’s going to cause huge problems, paragraph 12.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2024-10-17.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, When it comes to security, LLMs are like Swiss cheese — and that’s going to cause huge problems
Meanwhile, companies like Google, Tesla, and Figure.AI are now stuffing jailbreak-vulnerable LLMs into robots.
- Where
- paragraph 12
- Published
- 2024-10-17
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.