Embed Notice

HTML Code

<blockquote style="position: relative; padding-left: 55px;"><section><a href="https://thepit.social/users/peter/statuses/116684735163462360">Peter (peter@thepit.social)'s status on Wednesday, 03-Jun-2026 15:34:16 JST</a><a href="https://thepit.social/@peter" title="peter@thepit.social"><img src="https://gnusocial.jp/avatar/16909-48-20221104202315.webp" width="48" height="48" alt="Peter" style="position: absolute; left: 0; top: 0;">Peter</a><div><a href="https://thepit.social/@peter/116684693997496990" rel="in-reply-to">in reply to</a></div></section><article><p>uh-oh!! so this is something i heard Gary Marcus talking about on a podcast. the real recent progress in "AI" has been on the deterministic computing components **around** the LLMs, not the models themselves. building a good harness (think Claude Code) unlocks possibilities.</p></article><footer><a rel="bookmark" href="https://gnusocial.jp/conversation/6448909#notice-12695227">In conversation</a><time datetime="2026-06-03T15:34:16+09:00" title="Wednesday, 03-Jun-2026 15:34:16 JST">about 14 hours ago</time> <span>from <span><a href="https://thepit.social/@peter/116684735163462360" rel="external" title="Sent from thepit.social via ActivityPub">thepit.social</a></span></span><a href="https://thepit.social/@peter/116684735163462360">permalink</a><h4>Attachments</h4><ol><li><label><a rel="external" href="https://gnusocial.jp/attachment/6139534">We challenge previous findings. We distinguish information available to an LLM, which it acquires in the form of
knowledge extracted from the pretraining data and context extracted from the environment, from its ability to reason
and decide. Our results suggest that single-GPU models already possess sufficient decision-making capabilities to pose
severe cybersecurity risks. We hypothesize that their limited performance in past evaluations are primarily due to their
lack of a strong informational component, as their pretrained weights hold less knowledge. We show that a systematic
agentic harness can compensate for this gap to a surprising degree, by feeding the model targeted contextual information.
The informational component encompasses the extensive technical facts and exploit syntaxes encoded within the
parameters of larger models. For example, a smaller model lacking it might correctly deduce that a web server is</a></label><br><a href="https://cdn.masto.host/thepitsocial/media_attachments/files/116/684/720/898/097/748/original/8520f8faa7eb5f7d.png" rel="external">https://cdn.masto.host/thepitsocial/media_attachments/files/116/684/720/898/097/748/original/8520f8faa7eb5f7d.png</a></li><li><label><a rel="external" href="https://gnusocial.jp/attachment/6139535">vulnerable to an SQL injection based on reconnaissance, but fail to execute the attack because it cannot generate the exact
payload syntax zero-shot. The harness bridges this gap by feeding the model targeted contextual information, such as
the specific exploit or encoding syntax. Such injected contexts not only directly support decision-making but also unlock
latent pretrained knowledge, serving as a cue that enables the model to correctly reason through the rest of the attack.
Concretely, we demonstrate that, provided with the right informational support, a single-GPU LLM has sufficient
reasoning capabilities to generate attack strategies that enable the agent to penetrate victim machines: first by obtaining
initial command execution, a foothold, and then by escalating privileges to full administrative control. The agent then
leverages this control to replicate: it stages a copy of itself on the compromised machine, resolves the required runtime
dependencies, and launches an independent agent instance that discovers and attacks further targets.</a></label><br><a href="https://cdn.masto.host/thepitsocial/media_attachments/files/116/684/724/339/120/047/original/dbbfa679ca2f5b97.png" rel="external">https://cdn.masto.host/thepitsocial/media_attachments/files/116/684/724/339/120/047/original/dbbfa679ca2f5b97.png</a></li></ol></footer></blockquote>

Corresponding Notice

Embed this notice
Peter (peter@thepit.social)'s status on Wednesday, 03-Jun-2026 15:34:16 JST Peter
in reply to
uh-oh!! so this is something i heard Gary Marcus talking about on a podcast. the real recent progress in "AI" has been on the deterministic computing components **around** the LLMs, not the models themselves. building a good harness (think Claude Code) unlocks possibilities.
In conversationabout 14 hours ago from thepit.socialpermalink
Attachments
1. We challenge previous findings. We distinguish information available to an LLM, which it acquires in the form of knowledge extracted from the pretraining data and context extracted from the environment, from its ability to reason and decide. Our results suggest that single-GPU models already possess sufficient decision-making capabilities to pose severe cybersecurity risks. We hypothesize that their limited performance in past evaluations are primarily due to their lack of a strong informational component, as their pretrained weights hold less knowledge. We show that a systematic agentic harness can compensate for this gap to a surprising degree, by feeding the model targeted contextual information. The informational component encompasses the extensive technical facts and exploit syntaxes encoded within the parameters of larger models. For example, a smaller model lacking it might correctly deduce that a web server is
  https://cdn.masto.host/thepitsocial/media_attachments/files/116/684/720/898/097/748/original/8520f8faa7eb5f7d.png
2. vulnerable to an SQL injection based on reconnaissance, but fail to execute the attack because it cannot generate the exact payload syntax zero-shot. The harness bridges this gap by feeding the model targeted contextual information, such as the specific exploit or encoding syntax. Such injected contexts not only directly support decision-making but also unlock latent pretrained knowledge, serving as a cue that enables the model to correctly reason through the rest of the attack. Concretely, we demonstrate that, provided with the right informational support, a single-GPU LLM has sufficient reasoning capabilities to generate attack strategies that enable the agent to penetrate victim machines: first by obtaining initial command execution, a foothold, and then by escalating privileges to full administrative control. The agent then leverages this control to replicate: it stages a copy of itself on the compromised machine, resolves the required runtime dependencies, and launches an independent agent instance that discovers and attacks further targets.
  https://cdn.masto.host/thepitsocial/media_attachments/files/116/684/724/339/120/047/original/dbbfa679ca2f5b97.png

Public

Embed Notice

HTML Code

Corresponding Notice