Meta’s Muse Model Exposes File System Access in Unexpected Test

News at a Glance

Muse will now gladly zip up its entire filesystem for you, which Meta says is the “intended behavior.”

What Happened

On September 25, 2026, Meta officially unveiled Muse, a new foundation model positioned as a multimodal interaction assistant supporting text, image, and code generation. Within hours of release, multiple test users discovered that Muse, when answering specific queries, would proactively display file directory structures from its runtime environment, including some system configuration file paths. Meta subsequently issued a statement confirming the phenomenon stemmed from overgeneralization of tool-call data during training, and pushed an emergency patch to restrict file system-related instructions. As of September 26, Meta has suspended public API access to Muse and form

Background

This incident occurs during a period of heightened sensitivity to AI model safety boundaries. Since 2025, companies like OpenAI and Google have released models with tool-calling capabilities, but all have implemented strict sandbox mechanisms to limit system-level access. Meta's previous generation model, Llama 4, released in December 2025, introduced tool calling but did not open file system permissions. Muse, as Meta's first model deeply integrated with operating system-level interfaces, included a large volume of real tool-call logs in its training da

Deep Dive

Liu Gong believes the core value of the Muse incident lies in revealing the complexity of managing AI model capability boundaries. From a technical dimension, the model's generalization of tool calls far exceeds expectations, with tool-call logs in training data being internalized as executable behaviors rather than merely semantic references. Liu Gong points out this is not a simple prompt injection vulnerability but a behavioral fixation at the model weight level, which traditional rule-based safety filters cannot fundamentally address. From a commercial dimension, Meta's release pace has clearly accelerated, with Muse taking only 14 months from inception to launch compared to Llama 4's 22 months, leading to insufficient safety testing coverage. Liu Gong predicts that in the short term Meta will face declining developer trust, but in the long term this incident may push the industry toward stricter model behavior audit standards. From an industry impact dimension, the Muse incident parallels Anthropic's Claude tool-call overreach in 2025, where reinforcement learning reduced violation rates to 0.01%, while Meta has not yet published comparable data. Liu Gong believes this marks a shift in AI safety competition from defensive patches to systematic behavioral specification design.

Data Points

  • Muse has 1.8 trillion parameters, a 50% increase over Llama 4's 1.2 trillion; training data of 12 trillion tokens, up 50% from Llama 4's 8 trillion. Muse registered 500,000 users on launch day, compared to 350,000 for Llama 4 in the same period. Within 24 hours of the incident, Meta's stock fell 2.3%, and negative discussions about Muse in developer communities rose from 12% to 47%. In comparison, Anthropic's Claude achieved a tool-call violation rate of 0.01% after reinforcement learning, while Meta has not yet disclosed a baseline violation rate for Mu

Impact & Outlook

For the industry, in the short term of 3 months, leading AI companies may tighten model tool-call permissions and accelerate the release of security audit tools; in the medium term of 1 year, industry standards for model behavior traceability are expected to emerge. For ordinary users, in the short term of 3 months, public access to Muse will be restricted pending patched re-release; in the medium term of 1 year, users will see more safety disclosures before model use, though daily experiences may be slightly simplified due to permission tightening.

Reactions

Meta officially stated it has urgently patched and suspended API access, emphasizing user data was unaffected. An OpenAI spokesperson said it will strengthen sandbox testing for its own models' tool calls. Anthropic's research team noted on social media that the Muse incident validates their earlier research on risks of tool-call logs in training data. Several developer community users expressed reduced trust in Muse but look forward to re-evaluation after the patch.

Angles to Explore

Points to consider

  • How tool-call logs in training data affect behavioral generalization, and whether data desensitization mechanisms are needed
  • The balance between Meta's release pace and safety testing investment, reflecting industry competitive pressure
  • Whether regulatory frameworks for AI model system-level access should be subject to government review

Source & Original

This item comes from The Verge AI (published 2026-09-25 16:49:53). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.


Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.

Republishing: feel free to quote or repost with a link back to the original article on Wenfeng Radar. · Business & partnership: see our media kit
Daily briefing + Sunday review · Unsubscribe anytime by replying · Your email stays private
中文EN
Contact · Privacy · Disclaimer · AI Disclosure · Weekly Review · Media & Business
Email: wenfengleida@wenfengleida.cn